diff options
Diffstat (limited to 'src/gallium/drivers/llvmpipe')
78 files changed, 5014 insertions, 2779 deletions
diff --git a/src/gallium/drivers/llvmpipe/.gitignore b/src/gallium/drivers/llvmpipe/.gitignore new file mode 100644 index 0000000000..257b72d7b2 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/.gitignore @@ -0,0 +1 @@ +lp_tile_soa.c diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile index 5ac09de79e..e038a5229e 100644 --- a/src/gallium/drivers/llvmpipe/Makefile +++ b/src/gallium/drivers/llvmpipe/Makefile @@ -17,9 +17,14 @@ C_SOURCES = \ lp_bld_depth.c \ lp_bld_flow.c \ lp_bld_format_aos.c \ + lp_bld_format_query.c \ + lp_bld_format_soa.c \ lp_bld_interp.c \ lp_bld_intr.c \ lp_bld_logic.c \ + lp_bld_pack.c \ + lp_bld_sample.c \ + lp_bld_sample_soa.c \ lp_bld_swizzle.c \ lp_bld_struct.c \ lp_bld_tgsi_soa.c \ @@ -30,7 +35,6 @@ C_SOURCES = \ lp_draw_arrays.c \ lp_flush.c \ lp_jit.c \ - lp_prim_setup.c \ lp_prim_vbuf.c \ lp_setup.c \ lp_query.c \ @@ -46,9 +50,16 @@ C_SOURCES = \ lp_state_vs.c \ lp_surface.c \ lp_tex_cache.c \ - lp_tex_sample.c \ + lp_tex_sample_c.c \ + lp_tex_sample_llvm.c \ lp_texture.c \ lp_tile_cache.c \ lp_tile_soa.c +CPP_SOURCES = \ + lp_bld_misc.cpp + include ../../Makefile.template + +lp_tile_soa.c: lp_tile_soa.py ../../auxiliary/util/u_format_parse.py ../../auxiliary/util/u_format_access.py ../../auxiliary/util/u_format.csv + python lp_tile_soa.py ../../auxiliary/util/u_format.csv > $@ diff --git a/src/gallium/drivers/llvmpipe/README b/src/gallium/drivers/llvmpipe/README index 498d21dea6..89d08834a3 100644 --- a/src/gallium/drivers/llvmpipe/README +++ b/src/gallium/drivers/llvmpipe/README @@ -8,13 +8,16 @@ Done so far is: - the whole fragment pipeline is code generated in a single function + - input interpolation + - depth testing + - texture sampling (not all state/formats are supported) + - fragment shader TGSI translation - same level of support as the TGSI SSE2 exec machine, with the exception we don't fallback to TGSI interpretation when an unsupported opcode is found, but just ignore it - - texture sampling via an intrinsic call - done in SoA layout - input interpolation also code generated @@ -28,16 +31,17 @@ Done so far is: any width and length - not all operations are implemented for these types yet though -Most mesa/progs/demos/* work. Speed is on par with Keith's softpipe-opt branch, -which includes hand written fast implementations for common cases. +Most mesa/progs/demos/* work. To do (probably by this order): - code generate stipple and stencil testing - - code generate texture sampling + - translate the remaining bits of texture sampling state - translate TGSI control flow instructions, and all other remaining opcodes + + - integrate with the draw module for VS code generation - code generate the triangle setup and rasterization @@ -93,7 +97,7 @@ Alternatively, you can build it with GNU make, if you prefer, by invoking it as make linux-llvm -but the rest of these instructions assume scons is used. +but the rest of these instructions assume that scons is used. Using @@ -108,6 +112,9 @@ or export LD_LIBRARY_PATH=$PWD/build/linux-x86-debug/lib:$LD_LIBRARY_PATH +For performance evaluation pass debug=no to scons, and use the corresponding +lib directory without the "-debug" suffix. + Unit testing ============ @@ -119,7 +126,7 @@ build/linux-???-debug/gallium/drivers/llvmpipe: - lp_test_conv: SIMD vector conversion - lp_test_format: pixel unpacking/packing -Some of this tests can output results and benchmarks to a tab-seperated-file +Some of this tests can output results and benchmarks to a tab-separated-file for posterior analysis, e.g.: build/linux-x86_64-debug/gallium/drivers/llvmpipe/lp_test_blend -o blend.tsv @@ -133,10 +140,10 @@ Development Notes at the top of the lp_bld_*.c functions. - All lp_bld_*.[ch] are isolated from the rest of the driver, and could/may be - put in a standalone Gallium state -> LLVM IR translation module. + put in a stand-alone Gallium state -> LLVM IR translation module. - We use LLVM-C bindings for now. They are not documented, but follow the C++ interfaces very closely, and appear to be complete enough for code generation. See http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html - for a standalone example. + for a stand-alone example. diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript index 5c29bdac56..3bd2e70013 100644 --- a/src/gallium/drivers/llvmpipe/SConscript +++ b/src/gallium/drivers/llvmpipe/SConscript @@ -3,12 +3,19 @@ Import('*') env = env.Clone() env.Tool('llvm') -if 'LLVM_VERSION' not in env: +if not env.has_key('LLVM_VERSION'): print 'warning: LLVM not found: not building llvmpipe' Return() env.Tool('udis86') +env.CodeGenerate( + target = 'lp_tile_soa.c', + script = 'lp_tile_soa.py', + source = ['#src/gallium/auxiliary/util/u_format.csv'], + command = 'python $SCRIPT $SOURCE > $TARGET' +) + llvmpipe = env.ConvenienceLibrary( target = 'llvmpipe', source = [ @@ -23,8 +30,14 @@ llvmpipe = env.ConvenienceLibrary( 'lp_bld_depth.c', 'lp_bld_flow.c', 'lp_bld_format_aos.c', + 'lp_bld_format_query.c', + 'lp_bld_format_soa.c', 'lp_bld_interp.c', 'lp_bld_intr.c', + 'lp_bld_misc.cpp', + 'lp_bld_pack.c', + 'lp_bld_sample.c', + 'lp_bld_sample_soa.c', 'lp_bld_struct.c', 'lp_bld_logic.c', 'lp_bld_swizzle.c', @@ -36,7 +49,6 @@ llvmpipe = env.ConvenienceLibrary( 'lp_draw_arrays.c', 'lp_flush.c', 'lp_jit.c', - 'lp_prim_setup.c', 'lp_prim_vbuf.c', 'lp_setup.c', 'lp_query.c', @@ -52,7 +64,8 @@ llvmpipe = env.ConvenienceLibrary( 'lp_state_vs.c', 'lp_surface.c', 'lp_tex_cache.c', - 'lp_tex_sample.c', + 'lp_tex_sample_c.c', + 'lp_tex_sample_llvm.c', 'lp_texture.c', 'lp_tile_cache.c', 'lp_tile_soa.c', @@ -65,7 +78,7 @@ env.Prepend(LIBS = [llvmpipe] + auxiliaries) env.Program( target = 'lp_test_format', - source = ['lp_test_format.c'], + source = ['lp_test_format.c', 'lp_test_main.c'], ) env.Program( diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c index 49c2f911af..2b4bc5c819 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c @@ -45,7 +45,7 @@ void lp_build_alpha_test(LLVMBuilderRef builder, const struct pipe_alpha_state *state, - union lp_type type, + struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef alpha, LLVMValueRef ref) diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h index 9dbcdb4daa..634575670d 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h @@ -38,14 +38,14 @@ #include <llvm-c/Core.h> struct pipe_alpha_state; -union lp_type; +struct lp_type; struct lp_build_mask_context; void lp_build_alpha_test(LLVMBuilderRef builder, const struct pipe_alpha_state *state, - union lp_type type, + struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef alpha, LLVMValueRef ref); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c index 09a57ff33d..9c59677a74 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_arit.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c @@ -47,12 +47,16 @@ #include "util/u_memory.h" #include "util/u_debug.h" +#include "util/u_math.h" #include "util/u_string.h" +#include "util/u_cpu_detect.h" #include "lp_bld_type.h" #include "lp_bld_const.h" #include "lp_bld_intr.h" #include "lp_bld_logic.h" +#include "lp_bld_pack.h" +#include "lp_bld_debug.h" #include "lp_bld_arit.h" @@ -65,36 +69,34 @@ lp_build_min_simple(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; const char *intrinsic = NULL; LLVMValueRef cond; /* TODO: optimize the constant case */ -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) if(type.width * type.length == 128) { if(type.floating) { - if(type.width == 32) + if(type.width == 32 && util_cpu_caps.has_sse) intrinsic = "llvm.x86.sse.min.ps"; - if(type.width == 64) + if(type.width == 64 && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.min.pd"; } else { - if(type.width == 8 && !type.sign) + if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.pminu.b"; - if(type.width == 8 && type.sign) + if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pminsb"; - if(type.width == 16 && !type.sign) + if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pminuw"; - if(type.width == 16 && type.sign) + if(type.width == 16 && type.sign && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.pmins.w"; - if(type.width == 32 && !type.sign) + if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pminud"; - if(type.width == 32 && type.sign) + if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pminsd"; } } -#endif if(intrinsic) return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); @@ -113,36 +115,34 @@ lp_build_max_simple(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; const char *intrinsic = NULL; LLVMValueRef cond; /* TODO: optimize the constant case */ -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) if(type.width * type.length == 128) { if(type.floating) { - if(type.width == 32) + if(type.width == 32 && util_cpu_caps.has_sse) intrinsic = "llvm.x86.sse.max.ps"; - if(type.width == 64) + if(type.width == 64 && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.max.pd"; } else { - if(type.width == 8 && !type.sign) + if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.pmaxu.b"; - if(type.width == 8 && type.sign) + if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pmaxsb"; - if(type.width == 16 && !type.sign) + if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pmaxuw"; - if(type.width == 16 && type.sign) + if(type.width == 16 && type.sign && util_cpu_caps.has_sse2) intrinsic = "llvm.x86.sse2.pmaxs.w"; - if(type.width == 32 && !type.sign) + if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pmaxud"; - if(type.width == 32 && type.sign) + if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1) intrinsic = "llvm.x86.sse41.pmaxsd"; } } -#endif if(intrinsic) return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); @@ -159,7 +159,7 @@ LLVMValueRef lp_build_comp(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; if(a == bld->one) return bld->zero; @@ -188,7 +188,7 @@ lp_build_add(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMValueRef res; if(a == bld->zero) @@ -204,15 +204,14 @@ lp_build_add(struct lp_build_context *bld, if(a == bld->one || b == bld->one) return bld->one; -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width * type.length == 128 && + if(util_cpu_caps.has_sse2 && + type.width * type.length == 128 && !type.floating && !type.fixed) { if(type.width == 8) intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b"; if(type.width == 16) intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w"; } -#endif if(intrinsic) return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); @@ -241,7 +240,7 @@ lp_build_sub(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMValueRef res; if(b == bld->zero) @@ -257,15 +256,14 @@ lp_build_sub(struct lp_build_context *bld, if(b == bld->one) return bld->zero; -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width * type.length == 128 && + if(util_cpu_caps.has_sse2 && + type.width * type.length == 128 && !type.floating && !type.fixed) { if(type.width == 8) intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b"; if(type.width == 16) intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w"; } -#endif if(intrinsic) return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); @@ -284,45 +282,6 @@ lp_build_sub(struct lp_build_context *bld, /** - * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions. - */ -static LLVMValueRef -lp_build_unpack_shuffle(unsigned n, unsigned lo_hi) -{ - LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; - unsigned i, j; - - assert(n <= LP_MAX_VECTOR_LENGTH); - assert(lo_hi < 2); - - for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) { - elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0); - elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0); - } - - return LLVMConstVector(elems, n); -} - - -/** - * Build constant int vector of width 'n' and value 'c'. - */ -static LLVMValueRef -lp_build_const_vec(LLVMTypeRef type, unsigned n, long long c) -{ - LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; - unsigned i; - - assert(n <= LP_MAX_VECTOR_LENGTH); - - for(i = 0; i < n; ++i) - elems[i] = LLVMConstInt(type, c, 0); - - return LLVMConstVector(elems, n); -} - - -/** * Normalized 8bit multiplication. * * - alpha plus one @@ -365,33 +324,30 @@ lp_build_const_vec(LLVMTypeRef type, unsigned n, long long c) */ static LLVMValueRef lp_build_mul_u8n(LLVMBuilderRef builder, + struct lp_type i16_type, LLVMValueRef a, LLVMValueRef b) { - static LLVMValueRef c01 = NULL; - static LLVMValueRef c08 = NULL; - static LLVMValueRef c80 = NULL; + LLVMValueRef c8; LLVMValueRef ab; - if(!c01) c01 = lp_build_const_vec(LLVMInt16Type(), 8, 0x01); - if(!c08) c08 = lp_build_const_vec(LLVMInt16Type(), 8, 0x08); - if(!c80) c80 = lp_build_const_vec(LLVMInt16Type(), 8, 0x80); + c8 = lp_build_int_const_scalar(i16_type, 8); #if 0 /* a*b/255 ~= (a*(b + 1)) >> 256 */ - b = LLVMBuildAdd(builder, b, c01, ""); + b = LLVMBuildAdd(builder, b, lp_build_int_const_scalar(i16_type, 1), ""); ab = LLVMBuildMul(builder, a, b, ""); #else - /* t/255 ~= (t + (t >> 8) + 0x80) >> 8 */ + /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */ ab = LLVMBuildMul(builder, a, b, ""); - ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c08, ""), ""); - ab = LLVMBuildAdd(builder, ab, c80, ""); + ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c8, ""), ""); + ab = LLVMBuildAdd(builder, ab, lp_build_int_const_scalar(i16_type, 0x80), ""); #endif - ab = LLVMBuildLShr(builder, ab, c08, ""); + ab = LLVMBuildLShr(builder, ab, c8, ""); return ab; } @@ -405,7 +361,9 @@ lp_build_mul(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; + LLVMValueRef shift; + LLVMValueRef res; if(a == bld->zero) return bld->zero; @@ -419,53 +377,104 @@ lp_build_mul(struct lp_build_context *bld, return bld->undef; if(!type.floating && !type.fixed && type.norm) { -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width == 8 && type.length == 16) { - LLVMTypeRef i16x8 = LLVMVectorType(LLVMInt16Type(), 8); - LLVMTypeRef i8x16 = LLVMVectorType(LLVMInt8Type(), 16); - static LLVMValueRef ml = NULL; - static LLVMValueRef mh = NULL; - LLVMValueRef al, ah, bl, bh; - LLVMValueRef abl, abh; - LLVMValueRef ab; - - if(!ml) ml = lp_build_unpack_shuffle(16, 0); - if(!mh) mh = lp_build_unpack_shuffle(16, 1); - - /* PUNPCKLBW, PUNPCKHBW */ - al = LLVMBuildShuffleVector(bld->builder, a, bld->zero, ml, ""); - bl = LLVMBuildShuffleVector(bld->builder, b, bld->zero, ml, ""); - ah = LLVMBuildShuffleVector(bld->builder, a, bld->zero, mh, ""); - bh = LLVMBuildShuffleVector(bld->builder, b, bld->zero, mh, ""); + if(type.width == 8) { + struct lp_type i16_type = lp_wider_type(type); + LLVMValueRef al, ah, bl, bh, abl, abh, ab; - /* NOP */ - al = LLVMBuildBitCast(bld->builder, al, i16x8, ""); - bl = LLVMBuildBitCast(bld->builder, bl, i16x8, ""); - ah = LLVMBuildBitCast(bld->builder, ah, i16x8, ""); - bh = LLVMBuildBitCast(bld->builder, bh, i16x8, ""); + lp_build_unpack2(bld->builder, type, i16_type, a, &al, &ah); + lp_build_unpack2(bld->builder, type, i16_type, b, &bl, &bh); /* PMULLW, PSRLW, PADDW */ - abl = lp_build_mul_u8n(bld->builder, al, bl); - abh = lp_build_mul_u8n(bld->builder, ah, bh); + abl = lp_build_mul_u8n(bld->builder, i16_type, al, bl); + abh = lp_build_mul_u8n(bld->builder, i16_type, ah, bh); - /* PACKUSWB */ - ab = lp_build_intrinsic_binary(bld->builder, "llvm.x86.sse2.packuswb.128" , i16x8, abl, abh); - - /* NOP */ - ab = LLVMBuildBitCast(bld->builder, ab, i8x16, ""); + ab = lp_build_pack2(bld->builder, i16_type, type, abl, abh); return ab; } -#endif /* FIXME */ assert(0); } - if(LLVMIsConstant(a) && LLVMIsConstant(b)) - return LLVMConstMul(a, b); + if(type.fixed) + shift = lp_build_int_const_scalar(type, type.width/2); + else + shift = NULL; + + if(LLVMIsConstant(a) && LLVMIsConstant(b)) { + res = LLVMConstMul(a, b); + if(shift) { + if(type.sign) + res = LLVMConstAShr(res, shift); + else + res = LLVMConstLShr(res, shift); + } + } + else { + res = LLVMBuildMul(bld->builder, a, b, ""); + if(shift) { + if(type.sign) + res = LLVMBuildAShr(bld->builder, res, shift, ""); + else + res = LLVMBuildLShr(bld->builder, res, shift, ""); + } + } + + return res; +} + + +/** + * Small vector x scale multiplication optimization. + */ +LLVMValueRef +lp_build_mul_imm(struct lp_build_context *bld, + LLVMValueRef a, + int b) +{ + LLVMValueRef factor; + + if(b == 0) + return bld->zero; + + if(b == 1) + return a; + + if(b == -1) + return LLVMBuildNeg(bld->builder, a, ""); - return LLVMBuildMul(bld->builder, a, b, ""); + if(b == 2 && bld->type.floating) + return lp_build_add(bld, a, a); + + if(util_is_pot(b)) { + unsigned shift = ffs(b) - 1; + + if(bld->type.floating) { +#if 0 + /* + * Power of two multiplication by directly manipulating the mantissa. + * + * XXX: This might not be always faster, it will introduce a small error + * for multiplication by zero, and it will produce wrong results + * for Inf and NaN. + */ + unsigned mantissa = lp_mantissa(bld->type); + factor = lp_build_int_const_scalar(bld->type, (unsigned long long)shift << mantissa); + a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), ""); + a = LLVMBuildAdd(bld->builder, a, factor, ""); + a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), ""); + return a; +#endif + } + else { + factor = lp_build_const_scalar(bld->type, shift); + return LLVMBuildShl(bld->builder, a, factor, ""); + } + } + + factor = lp_build_const_scalar(bld->type, (double)b); + return lp_build_mul(bld, a, factor); } @@ -477,7 +486,7 @@ lp_build_div(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; if(a == bld->zero) return bld->zero; @@ -493,16 +502,62 @@ lp_build_div(struct lp_build_context *bld, if(LLVMIsConstant(a) && LLVMIsConstant(b)) return LLVMConstFDiv(a, b); -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width == 32 && type.length == 4) + if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) return lp_build_mul(bld, a, lp_build_rcp(bld, b)); -#endif return LLVMBuildFDiv(bld->builder, a, b, ""); } /** + * Linear interpolation. + * + * This also works for integer values with a few caveats. + * + * @sa http://www.stereopsis.com/doubleblend.html + */ +LLVMValueRef +lp_build_lerp(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef v0, + LLVMValueRef v1) +{ + LLVMValueRef delta; + LLVMValueRef res; + + delta = lp_build_sub(bld, v1, v0); + + res = lp_build_mul(bld, x, delta); + + res = lp_build_add(bld, v0, res); + + if(bld->type.fixed) + /* XXX: This step is necessary for lerping 8bit colors stored on 16bits, + * but it will be wrong for other uses. Basically we need a more + * powerful lp_type, capable of further distinguishing the values + * interpretation from the value storage. */ + res = LLVMBuildAnd(bld->builder, res, lp_build_int_const_scalar(bld->type, (1 << bld->type.width/2) - 1), ""); + + return res; +} + + +LLVMValueRef +lp_build_lerp_2d(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef v00, + LLVMValueRef v01, + LLVMValueRef v10, + LLVMValueRef v11) +{ + LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01); + LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11); + return lp_build_lerp(bld, y, v0, v1); +} + + +/** * Generate min(a, b) * Do checks for special cases. */ @@ -565,33 +620,333 @@ LLVMValueRef lp_build_abs(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); if(!type.sign) return a; - /* XXX: is this really necessary? */ -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(!type.floating && type.width*type.length == 128) { - LLVMTypeRef vec_type = lp_build_vec_type(type); - if(type.width == 8) + if(type.floating) { + /* Mask out the sign bit */ + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, ((unsigned long long)1 << type.width) - 1); + a = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + a = LLVMBuildAnd(bld->builder, a, mask, ""); + a = LLVMBuildBitCast(bld->builder, a, vec_type, ""); + return a; + } + + if(type.width*type.length == 128 && util_cpu_caps.has_ssse3) { + switch(type.width) { + case 8: return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a); - if(type.width == 16) + case 16: return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a); - if(type.width == 32) + case 32: return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a); + } } -#endif return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, "")); } LLVMValueRef +lp_build_sgn(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef cond; + LLVMValueRef res; + + /* Handle non-zero case */ + if(!type.sign) { + /* if not zero then sign must be positive */ + res = bld->one; + } + else if(type.floating) { + /* Take the sign bit and add it to 1 constant */ + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + LLVMValueRef sign; + LLVMValueRef one; + sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + sign = LLVMBuildAnd(bld->builder, sign, mask, ""); + one = LLVMConstBitCast(bld->one, int_vec_type); + res = LLVMBuildOr(bld->builder, sign, one, ""); + res = LLVMBuildBitCast(bld->builder, res, vec_type, ""); + } + else + { + LLVMValueRef minus_one = lp_build_const_scalar(type, -1.0); + cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero); + res = lp_build_select(bld, cond, bld->one, minus_one); + } + + /* Handle zero */ + cond = lp_build_cmp(bld, PIPE_FUNC_EQUAL, a, bld->zero); + res = lp_build_select(bld, cond, bld->zero, bld->one); + + return res; +} + + +enum lp_build_round_sse41_mode +{ + LP_BUILD_ROUND_SSE41_NEAREST = 0, + LP_BUILD_ROUND_SSE41_FLOOR = 1, + LP_BUILD_ROUND_SSE41_CEIL = 2, + LP_BUILD_ROUND_SSE41_TRUNCATE = 3 +}; + + +static INLINE LLVMValueRef +lp_build_round_sse41(struct lp_build_context *bld, + LLVMValueRef a, + enum lp_build_round_sse41_mode mode) +{ + const struct lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + const char *intrinsic; + + assert(type.floating); + assert(type.width*type.length == 128); + assert(lp_check_value(type, a)); + assert(util_cpu_caps.has_sse4_1); + + switch(type.width) { + case 32: + intrinsic = "llvm.x86.sse41.round.ps"; + break; + case 64: + intrinsic = "llvm.x86.sse41.round.pd"; + break; + default: + assert(0); + return bld->undef; + } + + return lp_build_intrinsic_binary(bld->builder, intrinsic, vec_type, a, + LLVMConstInt(LLVMInt32Type(), mode, 0)); +} + + +LLVMValueRef +lp_build_trunc(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) + return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE); + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, ""); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } +} + + +LLVMValueRef +lp_build_round(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) + return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST); + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_iround(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } +} + + +LLVMValueRef +lp_build_floor(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + + assert(type.floating); + + if(util_cpu_caps.has_sse4_1) + return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR); + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_ifloor(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } +} + + +LLVMValueRef +lp_build_ceil(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) + return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL); + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_iceil(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } +} + + +/** + * Convert to integer, through whichever rounding method that's fastest, + * typically truncating to zero. + */ +LLVMValueRef +lp_build_itrunc(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + + assert(type.floating); + assert(lp_check_value(type, a)); + + return LLVMBuildFPToSI(bld->builder, a, int_vec_type, ""); +} + + +LLVMValueRef +lp_build_iround(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST); + } + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + LLVMValueRef sign; + LLVMValueRef half; + + /* get sign bit */ + sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + sign = LLVMBuildAnd(bld->builder, sign, mask, ""); + + /* sign * 0.5 */ + half = lp_build_const_scalar(type, 0.5); + half = LLVMBuildBitCast(bld->builder, half, int_vec_type, ""); + half = LLVMBuildOr(bld->builder, sign, half, ""); + half = LLVMBuildBitCast(bld->builder, half, vec_type, ""); + + res = LLVMBuildAdd(bld->builder, a, half, ""); + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; +} + + +LLVMValueRef +lp_build_ifloor(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR); + } + else { + /* Take the sign bit and add it to 1 constant */ + LLVMTypeRef vec_type = lp_build_vec_type(type); + unsigned mantissa = lp_mantissa(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + LLVMValueRef sign; + LLVMValueRef offset; + + /* sign = a < 0 ? ~0 : 0 */ + sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + sign = LLVMBuildAnd(bld->builder, sign, mask, ""); + sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), ""); + + /* offset = -0.99999(9)f */ + offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa)); + offset = LLVMConstBitCast(offset, int_vec_type); + + /* offset = a < 0 ? -0.99999(9)f : 0.0f */ + offset = LLVMBuildAnd(bld->builder, offset, sign, ""); + offset = LLVMBuildBitCast(bld->builder, offset, vec_type, ""); + + res = LLVMBuildAdd(bld->builder, a, offset, ""); + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; +} + + +LLVMValueRef +lp_build_iceil(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL); + } + else { + assert(0); + res = bld->undef; + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; +} + + +LLVMValueRef lp_build_sqrt(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); char intrinsic[32]; @@ -609,7 +964,7 @@ LLVMValueRef lp_build_rcp(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; if(a == bld->zero) return bld->undef; @@ -623,11 +978,9 @@ lp_build_rcp(struct lp_build_context *bld, if(LLVMIsConstant(a)) return LLVMConstFDiv(bld->one, a); - /* XXX: is this really necessary? */ -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width == 32 && type.length == 4) + if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) + /* FIXME: improve precision */ return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a); -#endif return LLVMBuildFDiv(bld->builder, bld->one, a, ""); } @@ -640,15 +993,12 @@ LLVMValueRef lp_build_rsqrt(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; assert(type.floating); - /* XXX: is this really necessary? */ -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(type.width == 32 && type.length == 4) + if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type), a); -#endif return lp_build_rcp(bld, lp_build_sqrt(bld, a)); } @@ -661,7 +1011,7 @@ LLVMValueRef lp_build_cos(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); char intrinsic[32]; @@ -681,7 +1031,7 @@ LLVMValueRef lp_build_sin(struct lp_build_context *bld, LLVMValueRef a) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); char intrinsic[32]; @@ -704,7 +1054,8 @@ lp_build_pow(struct lp_build_context *bld, { /* TODO: optimize the constant case */ if(LLVMIsConstant(x) && LLVMIsConstant(y)) - debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + debug_printf("%s: inefficient/imprecise constant arithmetic\n", + __FUNCTION__); return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y)); } @@ -752,13 +1103,14 @@ lp_build_polynomial(struct lp_build_context *bld, const double *coeffs, unsigned num_coeffs) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMValueRef res = NULL; unsigned i; /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) - debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + debug_printf("%s: inefficient/imprecise constant arithmetic\n", + __FUNCTION__); for (i = num_coeffs; i--; ) { LLVMValueRef coeff = lp_build_const_scalar(type, coeffs[i]); @@ -800,7 +1152,7 @@ lp_build_exp2_approx(struct lp_build_context *bld, LLVMValueRef *p_frac_part, LLVMValueRef *p_exp2) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); LLVMValueRef ipart = NULL; @@ -812,7 +1164,8 @@ lp_build_exp2_approx(struct lp_build_context *bld, if(p_exp2_int_part || p_frac_part || p_exp2) { /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) - debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + debug_printf("%s: inefficient/imprecise constant arithmetic\n", + __FUNCTION__); assert(type.floating && type.width == 32); @@ -893,7 +1246,7 @@ lp_build_log2_approx(struct lp_build_context *bld, LLVMValueRef *p_floor_log2, LLVMValueRef *p_log2) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); @@ -911,7 +1264,8 @@ lp_build_log2_approx(struct lp_build_context *bld, if(p_exp || p_floor_log2 || p_log2) { /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) - debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + debug_printf("%s: inefficient/imprecise constant arithmetic\n", + __FUNCTION__); assert(type.floating && type.width == 32); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h index fc8cb25966..62be4b9aee 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_arit.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h @@ -40,7 +40,7 @@ #include <llvm-c/Core.h> -union lp_type type; +struct lp_type; struct lp_build_context; @@ -67,11 +67,36 @@ lp_build_mul(struct lp_build_context *bld, LLVMValueRef b); LLVMValueRef +lp_build_mul_imm(struct lp_build_context *bld, + LLVMValueRef a, + int b); + +LLVMValueRef lp_build_div(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b); LLVMValueRef +lp_build_lerp(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef v0, + LLVMValueRef v1); + +/** + * Bilinear interpolation. + * + * Values indices are in v_{yx}. + */ +LLVMValueRef +lp_build_lerp_2d(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef v00, + LLVMValueRef v01, + LLVMValueRef v10, + LLVMValueRef v11); + +LLVMValueRef lp_build_min(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b); @@ -86,6 +111,41 @@ lp_build_abs(struct lp_build_context *bld, LLVMValueRef a); LLVMValueRef +lp_build_sgn(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_round(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_floor(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_ceil(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_trunc(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_ifloor(struct lp_build_context *bld, + LLVMValueRef a); +LLVMValueRef +lp_build_iceil(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_iround(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_itrunc(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef lp_build_sqrt(struct lp_build_context *bld, LLVMValueRef a); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend.h b/src/gallium/drivers/llvmpipe/lp_bld_blend.h index d19e18846c..da272e549f 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_blend.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend.h @@ -46,7 +46,7 @@ struct pipe_blend_state; -union lp_type; +struct lp_type; struct lp_build_context; @@ -74,7 +74,7 @@ lp_build_blend_func(struct lp_build_context *bld, LLVMValueRef lp_build_blend_aos(LLVMBuilderRef builder, const struct pipe_blend_state *blend, - union lp_type type, + struct lp_type type, LLVMValueRef src, LLVMValueRef dst, LLVMValueRef const_, @@ -84,7 +84,7 @@ lp_build_blend_aos(LLVMBuilderRef builder, void lp_build_blend_soa(LLVMBuilderRef builder, const struct pipe_blend_state *blend, - union lp_type type, + struct lp_type type, LLVMValueRef src[4], LLVMValueRef dst[4], LLVMValueRef const_[4], diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c index c11a9398f8..d14f468ba9 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c @@ -303,7 +303,7 @@ lp_build_blend_func(struct lp_build_context *bld, LLVMValueRef lp_build_blend_aos(LLVMBuilderRef builder, const struct pipe_blend_state *blend, - union lp_type type, + struct lp_type type, LLVMValueRef src, LLVMValueRef dst, LLVMValueRef const_, diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c index b92254a7d6..9511299d55 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c @@ -199,7 +199,7 @@ lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld, void lp_build_blend_soa(LLVMBuilderRef builder, const struct pipe_blend_state *blend, - union lp_type type, + struct lp_type type, LLVMValueRef src[4], LLVMValueRef dst[4], LLVMValueRef con[4], diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.c b/src/gallium/drivers/llvmpipe/lp_bld_const.c index 21487365ea..c8eaa8c394 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_const.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_const.c @@ -42,7 +42,7 @@ unsigned -lp_mantissa(union lp_type type) +lp_mantissa(struct lp_type type) { assert(type.floating); @@ -72,7 +72,7 @@ lp_mantissa(union lp_type type) * Same as lp_const_scale(), but in terms of shifts. */ unsigned -lp_const_shift(union lp_type type) +lp_const_shift(struct lp_type type) { if(type.floating) return 0; @@ -86,7 +86,7 @@ lp_const_shift(union lp_type type) unsigned -lp_const_offset(union lp_type type) +lp_const_offset(struct lp_type type) { if(type.floating || type.fixed) return 0; @@ -104,7 +104,7 @@ lp_const_offset(union lp_type type) * else for the fixed points types and normalized integers. */ double -lp_const_scale(union lp_type type) +lp_const_scale(struct lp_type type) { unsigned long long llscale; double dscale; @@ -122,7 +122,7 @@ lp_const_scale(union lp_type type) * Minimum value representable by the type. */ double -lp_const_min(union lp_type type) +lp_const_min(struct lp_type type) { unsigned bits; @@ -158,7 +158,7 @@ lp_const_min(union lp_type type) * Maximum value representable by the type. */ double -lp_const_max(union lp_type type) +lp_const_max(struct lp_type type) { unsigned bits; @@ -190,7 +190,7 @@ lp_const_max(union lp_type type) double -lp_const_eps(union lp_type type) +lp_const_eps(struct lp_type type) { if (type.floating) { switch(type.width) { @@ -211,7 +211,7 @@ lp_const_eps(union lp_type type) LLVMValueRef -lp_build_undef(union lp_type type) +lp_build_undef(struct lp_type type) { LLVMTypeRef vec_type = lp_build_vec_type(type); return LLVMGetUndef(vec_type); @@ -219,7 +219,7 @@ lp_build_undef(union lp_type type) LLVMValueRef -lp_build_zero(union lp_type type) +lp_build_zero(struct lp_type type) { LLVMTypeRef vec_type = lp_build_vec_type(type); return LLVMConstNull(vec_type); @@ -227,7 +227,7 @@ lp_build_zero(union lp_type type) LLVMValueRef -lp_build_one(union lp_type type) +lp_build_one(struct lp_type type) { LLVMTypeRef elem_type; LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; @@ -269,7 +269,7 @@ lp_build_one(union lp_type type) LLVMValueRef -lp_build_const_scalar(union lp_type type, +lp_build_const_scalar(struct lp_type type, double val) { LLVMTypeRef elem_type = lp_build_elem_type(type); @@ -295,7 +295,7 @@ lp_build_const_scalar(union lp_type type, LLVMValueRef -lp_build_int_const_scalar(union lp_type type, +lp_build_int_const_scalar(struct lp_type type, long long val) { LLVMTypeRef elem_type = lp_build_int_elem_type(type); @@ -312,7 +312,7 @@ lp_build_int_const_scalar(union lp_type type, LLVMValueRef -lp_build_const_aos(union lp_type type, +lp_build_const_aos(struct lp_type type, double r, double g, double b, double a, const unsigned char *swizzle) { @@ -352,8 +352,8 @@ lp_build_const_aos(union lp_type type, LLVMValueRef -lp_build_const_mask_aos(union lp_type type, - boolean cond[4]) +lp_build_const_mask_aos(struct lp_type type, + const boolean cond[4]) { LLVMTypeRef elem_type = LLVMIntType(type.width); LLVMValueRef masks[LP_MAX_VECTOR_LENGTH]; diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.h b/src/gallium/drivers/llvmpipe/lp_bld_const.h index 1934530ea3..cb8e1c7b00 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_const.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_const.h @@ -42,67 +42,67 @@ #include <pipe/p_compiler.h> -union lp_type type; +struct lp_type; unsigned -lp_mantissa(union lp_type type); +lp_mantissa(struct lp_type type); unsigned -lp_const_shift(union lp_type type); +lp_const_shift(struct lp_type type); unsigned -lp_const_offset(union lp_type type); +lp_const_offset(struct lp_type type); double -lp_const_scale(union lp_type type); +lp_const_scale(struct lp_type type); double -lp_const_min(union lp_type type); +lp_const_min(struct lp_type type); double -lp_const_max(union lp_type type); +lp_const_max(struct lp_type type); double -lp_const_eps(union lp_type type); +lp_const_eps(struct lp_type type); LLVMValueRef -lp_build_undef(union lp_type type); +lp_build_undef(struct lp_type type); LLVMValueRef -lp_build_zero(union lp_type type); +lp_build_zero(struct lp_type type); LLVMValueRef -lp_build_one(union lp_type type); +lp_build_one(struct lp_type type); LLVMValueRef -lp_build_const_scalar(union lp_type type, +lp_build_const_scalar(struct lp_type type, double val); LLVMValueRef -lp_build_int_const_scalar(union lp_type type, +lp_build_int_const_scalar(struct lp_type type, long long val); LLVMValueRef -lp_build_const_aos(union lp_type type, +lp_build_const_aos(struct lp_type type, double r, double g, double b, double a, const unsigned char *swizzle); LLVMValueRef -lp_build_const_mask_aos(union lp_type type, - boolean cond[4]); +lp_build_const_mask_aos(struct lp_type type, + const boolean cond[4]); #endif /* !LP_BLD_CONST_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.c b/src/gallium/drivers/llvmpipe/lp_bld_conv.c index c8954c8a34..9935209437 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_conv.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.c @@ -63,11 +63,13 @@ #include "util/u_debug.h" #include "util/u_math.h" +#include "util/u_cpu_detect.h" #include "lp_bld_type.h" #include "lp_bld_const.h" #include "lp_bld_intr.h" #include "lp_bld_arit.h" +#include "lp_bld_pack.h" #include "lp_bld_conv.h" @@ -86,7 +88,7 @@ */ LLVMValueRef lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, - union lp_type src_type, + struct lp_type src_type, unsigned dst_width, LLVMValueRef src) { @@ -122,7 +124,7 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, int shift = dst_width - n; res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), ""); - /* Fill in the empty lower bits for added precision? */ + /* TODO: Fill in the empty lower bits for additional precision? */ #if 0 { LLVMValueRef msb; @@ -152,7 +154,7 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, LLVMValueRef lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, unsigned src_width, - union lp_type dst_type, + struct lp_type dst_type, LLVMValueRef src) { LLVMTypeRef vec_type = lp_build_vec_type(dst_type); @@ -198,243 +200,6 @@ lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, /** - * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions. - */ -static LLVMValueRef -lp_build_const_unpack_shuffle(unsigned n, unsigned lo_hi) -{ - LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; - unsigned i, j; - - assert(n <= LP_MAX_VECTOR_LENGTH); - assert(lo_hi < 2); - - /* TODO: cache results in a static table */ - - for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) { - elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0); - elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0); - } - - return LLVMConstVector(elems, n); -} - - -/** - * Build shuffle vectors that match PACKxx instructions. - */ -static LLVMValueRef -lp_build_const_pack_shuffle(unsigned n) -{ - LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; - unsigned i; - - assert(n <= LP_MAX_VECTOR_LENGTH); - - /* TODO: cache results in a static table */ - - for(i = 0; i < n; ++i) - elems[i] = LLVMConstInt(LLVMInt32Type(), 2*i, 0); - - return LLVMConstVector(elems, n); -} - - -/** - * Expand the bit width. - * - * This will only change the number of bits the values are represented, not the - * values themselved. - */ -static void -lp_build_expand(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, - LLVMValueRef src, - LLVMValueRef *dst, unsigned num_dsts) -{ - unsigned num_tmps; - unsigned i; - - /* Register width must remain constant */ - assert(src_type.width * src_type.length == dst_type.width * dst_type.length); - - /* We must not loose or gain channels. Only precision */ - assert(src_type.length == dst_type.length * num_dsts); - - num_tmps = 1; - dst[0] = src; - - while(src_type.width < dst_type.width) { - union lp_type new_type = src_type; - LLVMTypeRef new_vec_type; - - new_type.width *= 2; - new_type.length /= 2; - new_vec_type = lp_build_vec_type(new_type); - - for(i = num_tmps; i--; ) { - LLVMValueRef zero; - LLVMValueRef shuffle_lo; - LLVMValueRef shuffle_hi; - LLVMValueRef lo; - LLVMValueRef hi; - - zero = lp_build_zero(src_type); - shuffle_lo = lp_build_const_unpack_shuffle(src_type.length, 0); - shuffle_hi = lp_build_const_unpack_shuffle(src_type.length, 1); - - /* PUNPCKLBW, PUNPCKHBW */ - lo = LLVMBuildShuffleVector(builder, dst[i], zero, shuffle_lo, ""); - hi = LLVMBuildShuffleVector(builder, dst[i], zero, shuffle_hi, ""); - - dst[2*i + 0] = LLVMBuildBitCast(builder, lo, new_vec_type, ""); - dst[2*i + 1] = LLVMBuildBitCast(builder, hi, new_vec_type, ""); - } - - src_type = new_type; - - num_tmps *= 2; - } - - assert(num_tmps == num_dsts); -} - - -/** - * Non-interleaved pack. - * - * This will move values as - * - * lo = __ l0 __ l1 __ l2 __.. __ ln - * hi = __ h0 __ h1 __ h2 __.. __ hn - * res = l0 l1 l2 .. ln h0 h1 h2 .. hn - * - * TODO: handle saturation consistently. - */ -static LLVMValueRef -lp_build_pack2(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, - boolean clamped, - LLVMValueRef lo, - LLVMValueRef hi) -{ - LLVMTypeRef src_vec_type = lp_build_vec_type(src_type); - LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type); - LLVMValueRef shuffle; - LLVMValueRef res; - - /* Register width must remain constant */ - assert(src_type.width * src_type.length == dst_type.width * dst_type.length); - - /* We must not loose or gain channels. Only precision */ - assert(src_type.length * 2 == dst_type.length); - - assert(!src_type.floating); - assert(!dst_type.floating); - -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) - if(src_type.width * src_type.length == 128) { - /* All X86 non-interleaved pack instructions all take signed inputs and - * saturate them, so saturate beforehand. */ - if(!src_type.sign && !clamped) { - struct lp_build_context bld; - unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width; - LLVMValueRef dst_max = lp_build_int_const_scalar(src_type, ((unsigned long long)1 << dst_bits) - 1); - lp_build_context_init(&bld, builder, src_type); - lo = lp_build_min(&bld, lo, dst_max); - hi = lp_build_min(&bld, hi, dst_max); - } - - switch(src_type.width) { - case 32: - if(dst_type.sign) - res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", src_vec_type, lo, hi); - else - /* PACKUSDW is the only instrinsic with a consistent signature */ - return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi); - break; - - case 16: - if(dst_type.sign) - res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", src_vec_type, lo, hi); - else - res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", src_vec_type, lo, hi); - break; - - default: - assert(0); - return LLVMGetUndef(dst_vec_type); - break; - } - - res = LLVMBuildBitCast(builder, res, dst_vec_type, ""); - return res; - } -#endif - - lo = LLVMBuildBitCast(builder, lo, dst_vec_type, ""); - hi = LLVMBuildBitCast(builder, hi, dst_vec_type, ""); - - shuffle = lp_build_const_pack_shuffle(dst_type.length); - - res = LLVMBuildShuffleVector(builder, lo, hi, shuffle, ""); - - return res; -} - - -/** - * Truncate the bit width. - * - * TODO: Handle saturation consistently. - */ -static LLVMValueRef -lp_build_trunc(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, - boolean clamped, - const LLVMValueRef *src, unsigned num_srcs) -{ - LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH]; - unsigned i; - - /* Register width must remain constant */ - assert(src_type.width * src_type.length == dst_type.width * dst_type.length); - - /* We must not loose or gain channels. Only precision */ - assert(src_type.length * num_srcs == dst_type.length); - - for(i = 0; i < num_srcs; ++i) - tmp[i] = src[i]; - - while(src_type.width > dst_type.width) { - union lp_type new_type = src_type; - - new_type.width /= 2; - new_type.length *= 2; - - /* Take in consideration the sign changes only in the last step */ - if(new_type.width == dst_type.width) - new_type.sign = dst_type.sign; - - num_srcs /= 2; - - for(i = 0; i < num_srcs; ++i) - tmp[i] = lp_build_pack2(builder, src_type, new_type, clamped, - tmp[2*i + 0], tmp[2*i + 1]); - - src_type = new_type; - } - - assert(num_srcs == 1); - - return tmp[0]; -} - - -/** * Generic type conversion. * * TODO: Take a precision argument, or even better, add a new precision member @@ -442,12 +207,12 @@ lp_build_trunc(LLVMBuilderRef builder, */ void lp_build_conv(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, + struct lp_type src_type, + struct lp_type dst_type, const LLVMValueRef *src, unsigned num_srcs, LLVMValueRef *dst, unsigned num_dsts) { - union lp_type tmp_type; + struct lp_type tmp_type; LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH]; unsigned num_tmps; unsigned i; @@ -470,7 +235,7 @@ lp_build_conv(LLVMBuilderRef builder, * Clamp if necessary */ - if(src_type.value != dst_type.value) { + if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) { struct lp_build_context bld; double src_min = lp_const_min(src_type); double dst_min = lp_const_min(dst_type); @@ -565,7 +330,7 @@ lp_build_conv(LLVMBuilderRef builder, if(tmp_type.width > dst_type.width) { assert(num_dsts == 1); - tmp[0] = lp_build_trunc(builder, tmp_type, dst_type, TRUE, tmp, num_tmps); + tmp[0] = lp_build_pack(builder, tmp_type, dst_type, TRUE, tmp, num_tmps); tmp_type.width = dst_type.width; tmp_type.length = dst_type.length; num_tmps = 1; @@ -573,7 +338,7 @@ lp_build_conv(LLVMBuilderRef builder, if(tmp_type.width < dst_type.width) { assert(num_tmps == 1); - lp_build_expand(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts); + lp_build_unpack(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts); tmp_type.width = dst_type.width; tmp_type.length = dst_type.length; num_tmps = num_dsts; @@ -656,8 +421,8 @@ lp_build_conv(LLVMBuilderRef builder, */ void lp_build_conv_mask(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, + struct lp_type src_type, + struct lp_type dst_type, const LLVMValueRef *src, unsigned num_srcs, LLVMValueRef *dst, unsigned num_dsts) { @@ -689,11 +454,11 @@ lp_build_conv_mask(LLVMBuilderRef builder, if(src_type.width > dst_type.width) { assert(num_dsts == 1); - dst[0] = lp_build_trunc(builder, src_type, dst_type, TRUE, src, num_srcs); + dst[0] = lp_build_pack(builder, src_type, dst_type, TRUE, src, num_srcs); } else if(src_type.width < dst_type.width) { assert(num_srcs == 1); - lp_build_expand(builder, src_type, dst_type, src[0], dst, num_dsts); + lp_build_unpack(builder, src_type, dst_type, src[0], dst, num_dsts); } else { assert(num_srcs == num_dsts); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.h b/src/gallium/drivers/llvmpipe/lp_bld_conv.h index 05c1ef2a10..948e68fae4 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_conv.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.h @@ -40,33 +40,33 @@ #include <llvm-c/Core.h> -union lp_type type; +struct lp_type; LLVMValueRef lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, - union lp_type src_type, + struct lp_type src_type, unsigned dst_width, LLVMValueRef src); LLVMValueRef lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, unsigned src_width, - union lp_type dst_type, + struct lp_type dst_type, LLVMValueRef src); void lp_build_conv(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, + struct lp_type src_type, + struct lp_type dst_type, const LLVMValueRef *srcs, unsigned num_srcs, LLVMValueRef *dsts, unsigned num_dsts); void lp_build_conv_mask(LLVMBuilderRef builder, - union lp_type src_type, - union lp_type dst_type, + struct lp_type src_type, + struct lp_type dst_type, const LLVMValueRef *src, unsigned num_srcs, LLVMValueRef *dst, unsigned num_dsts); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.c b/src/gallium/drivers/llvmpipe/lp_bld_debug.c index 30925b5f41..59d8f492e6 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_debug.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.c @@ -30,10 +30,27 @@ #include <udis86.h> #endif +#include "util/u_math.h" #include "util/u_debug.h" #include "lp_bld_debug.h" +/** + * Check alignment. + * + * It is important that this check is not implemented as a macro or inlined + * function, as the compiler assumptions in respect to alignment of global + * and stack variables would often make the check a no op, defeating the + * whole purpose of the exercise. + */ +boolean +lp_check_alignment(const void *ptr, unsigned alignment) +{ + assert(util_is_pot(alignment)); + return ((uintptr_t)ptr & (alignment - 1)) == 0; +} + + void lp_disassemble(const void* func) { diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.h b/src/gallium/drivers/llvmpipe/lp_bld_debug.h index ecdafef76d..583e6132b4 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_debug.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.h @@ -53,6 +53,10 @@ lp_build_name(LLVMValueRef val, const char *format, ...) } +boolean +lp_check_alignment(const void *ptr, unsigned alignment); + + void lp_disassemble(const void* func); diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.c b/src/gallium/drivers/llvmpipe/lp_bld_depth.c index 2cd6e6b921..98ec1cb1b9 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_depth.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c @@ -71,11 +71,11 @@ /** * Return a type appropriate for depth/stencil testing. */ -union lp_type +struct lp_type lp_depth_type(const struct util_format_description *format_desc, unsigned length) { - union lp_type type; + struct lp_type type; unsigned swizzle; assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS); @@ -85,7 +85,7 @@ lp_depth_type(const struct util_format_description *format_desc, swizzle = format_desc->swizzle[0]; assert(swizzle < 4); - type.value = 0; + memset(&type, 0, sizeof type); type.width = format_desc->block.bits; if(format_desc->channel[swizzle].type == UTIL_FORMAT_TYPE_FLOAT) { @@ -114,7 +114,7 @@ lp_depth_type(const struct util_format_description *format_desc, void lp_build_depth_test(LLVMBuilderRef builder, const struct pipe_depth_state *state, - union lp_type type, + struct lp_type type, const struct util_format_description *format_desc, struct lp_build_mask_context *mask, LLVMValueRef src, @@ -179,12 +179,13 @@ lp_build_depth_test(LLVMBuilderRef builder, padding_right = 0; for(chan = 0; chan < z_swizzle; ++chan) padding_right += format_desc->channel[chan].size; - padding_left = format_desc->block.bits - format_desc->channel[z_swizzle].size; + padding_left = format_desc->block.bits - + (padding_right + format_desc->channel[z_swizzle].size); if(padding_left || padding_right) { - const long long mask_left = ((long long)1 << (format_desc->block.bits - padding_left)) - 1; - const long long mask_right = ((long long)1 << (padding_right)) - 1; - z_bitmask = lp_build_int_const_scalar(type, mask_left & mask_right); + const unsigned long long mask_left = ((unsigned long long)1 << (format_desc->block.bits - padding_left)) - 1; + const unsigned long long mask_right = ((unsigned long long)1 << (padding_right)) - 1; + z_bitmask = lp_build_int_const_scalar(type, mask_left ^ mask_right); } if(padding_left) @@ -209,6 +210,4 @@ lp_build_depth_test(LLVMBuilderRef builder, dst = lp_build_select(&bld, z_bitmask, src, dst); LLVMBuildStore(builder, dst, dst_ptr); } - - assert(!state->occlusion_count); } diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.h b/src/gallium/drivers/llvmpipe/lp_bld_depth.h index 5d2e042fcc..79d6981bb5 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_depth.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.h @@ -41,11 +41,11 @@ struct pipe_depth_state; struct util_format_description; -union lp_type; +struct lp_type; struct lp_build_mask_context; -union lp_type +struct lp_type lp_depth_type(const struct util_format_description *format_desc, unsigned length); @@ -53,7 +53,7 @@ lp_depth_type(const struct util_format_description *format_desc, void lp_build_depth_test(LLVMBuilderRef builder, const struct pipe_depth_state *state, - union lp_type type, + struct lp_type type, const struct util_format_description *format_desc, struct lp_build_mask_context *mask, LLVMValueRef src, diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.c b/src/gallium/drivers/llvmpipe/lp_bld_flow.c index 9d99e1a9d9..dcc25fbff8 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_flow.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.c @@ -32,59 +32,261 @@ */ #include "util/u_debug.h" +#include "util/u_memory.h" #include "lp_bld_type.h" #include "lp_bld_flow.h" +#define LP_BUILD_FLOW_MAX_VARIABLES 32 +#define LP_BUILD_FLOW_MAX_DEPTH 32 + + +/** + * Enumeration of all possible flow constructs. + */ +enum lp_build_flow_construct_kind { + lP_BUILD_FLOW_SCOPE, + LP_BUILD_FLOW_SKIP, +}; + + +/** + * Variable declaration scope. + */ +struct lp_build_flow_scope +{ + /** Number of variables declared in this scope */ + unsigned num_variables; +}; + + +/** + * Early exit. Useful to skip to the end of a function or block when + * the execution mask becomes zero or when there is an error condition. + */ +struct lp_build_flow_skip +{ + /** Block to skip to */ + LLVMBasicBlockRef block; + + /** Number of variables declared at the beginning */ + unsigned num_variables; + + LLVMValueRef *phi; +}; + + +/** + * Union of all possible flow constructs' data + */ +union lp_build_flow_construct_data +{ + struct lp_build_flow_scope scope; + struct lp_build_flow_skip skip; +}; + + +/** + * Element of the flow construct stack. + */ +struct lp_build_flow_construct +{ + enum lp_build_flow_construct_kind kind; + union lp_build_flow_construct_data data; +}; + + +/** + * All necessary data to generate LLVM control flow constructs. + * + * Besides keeping track of the control flow construct themselves we also + * need to keep track of variables in order to generate SSA Phi values. + */ +struct lp_build_flow_context +{ + LLVMBuilderRef builder; + + /** + * Control flow stack. + */ + struct lp_build_flow_construct constructs[LP_BUILD_FLOW_MAX_DEPTH]; + unsigned num_constructs; + + /** + * Variable stack + */ + LLVMValueRef *variables[LP_BUILD_FLOW_MAX_VARIABLES]; + unsigned num_variables; +}; + + +struct lp_build_flow_context * +lp_build_flow_create(LLVMBuilderRef builder) +{ + struct lp_build_flow_context *flow; + + flow = CALLOC_STRUCT(lp_build_flow_context); + if(!flow) + return NULL; + + flow->builder = builder; + + return flow; +} + + void -lp_build_mask_begin(struct lp_build_mask_context *mask, - LLVMBuilderRef builder, - union lp_type type, - LLVMValueRef value) +lp_build_flow_destroy(struct lp_build_flow_context *flow) { - memset(mask, 0, sizeof *mask); + assert(flow->num_constructs == 0); + assert(flow->num_variables == 0); + FREE(flow); +} - mask->builder = builder; - mask->reg_type = LLVMIntType(type.width * type.length); - mask->value = value; + +static union lp_build_flow_construct_data * +lp_build_flow_push(struct lp_build_flow_context *flow, + enum lp_build_flow_construct_kind kind) +{ + assert(flow->num_constructs < LP_BUILD_FLOW_MAX_DEPTH); + if(flow->num_constructs >= LP_BUILD_FLOW_MAX_DEPTH) + return NULL; + + flow->constructs[flow->num_constructs].kind = kind; + return &flow->constructs[flow->num_constructs++].data; +} + + +static union lp_build_flow_construct_data * +lp_build_flow_peek(struct lp_build_flow_context *flow, + enum lp_build_flow_construct_kind kind) +{ + assert(flow->num_constructs); + if(!flow->num_constructs) + return NULL; + + assert(flow->constructs[flow->num_constructs - 1].kind == kind); + if(flow->constructs[flow->num_constructs - 1].kind != kind) + return NULL; + + return &flow->constructs[flow->num_constructs - 1].data; } +static union lp_build_flow_construct_data * +lp_build_flow_pop(struct lp_build_flow_context *flow, + enum lp_build_flow_construct_kind kind) +{ + assert(flow->num_constructs); + if(!flow->num_constructs) + return NULL; + + assert(flow->constructs[flow->num_constructs - 1].kind == kind); + if(flow->constructs[flow->num_constructs - 1].kind != kind) + return NULL; + + return &flow->constructs[--flow->num_constructs].data; +} + + +/** + * Begin a variable scope. + * + * + */ void -lp_build_mask_update(struct lp_build_mask_context *mask, - LLVMValueRef value) +lp_build_flow_scope_begin(struct lp_build_flow_context *flow) { + struct lp_build_flow_scope *scope; - LLVMValueRef cond; - LLVMBasicBlockRef current_block; - LLVMBasicBlockRef next_block; - LLVMBasicBlockRef new_block; + scope = &lp_build_flow_push(flow, lP_BUILD_FLOW_SCOPE)->scope; + if(!scope) + return; - if(mask->value) - mask->value = LLVMBuildAnd(mask->builder, mask->value, value, ""); - else - mask->value = value; + scope->num_variables = 0; +} - /* FIXME: disabled until we have proper control flow helpers */ -#if 0 - cond = LLVMBuildICmp(mask->builder, - LLVMIntEQ, - LLVMBuildBitCast(mask->builder, mask->value, mask->reg_type, ""), - LLVMConstNull(mask->reg_type), - ""); - current_block = LLVMGetInsertBlock(mask->builder); +/** + * Declare a variable. + * + * A variable is a named entity which can have different LLVMValueRef's at + * different points of the program. This is relevant for control flow because + * when there are mutiple branches to a same location we need to replace + * the variable's value with a Phi function as explained in + * http://en.wikipedia.org/wiki/Static_single_assignment_form . + * + * We keep track of variables by keeping around a pointer to where their + * current. + * + * There are a few cautions to observe: + * + * - Variable's value must not be NULL. If there is no initial value then + * LLVMGetUndef() should be used. + * + * - Variable's value must be kept up-to-date. If the variable is going to be + * modified by a function then a pointer should be passed so that its value + * is accurate. Failure to do this will cause some of the variables' + * transient values to be lost, leading to wrong results. + * + * - A program should be written from top to bottom, by always appending + * instructions to the bottom with a single LLVMBuilderRef. Inserting and/or + * modifying existing statements will most likely lead to wrong results. + * + */ +void +lp_build_flow_scope_declare(struct lp_build_flow_context *flow, + LLVMValueRef *variable) +{ + struct lp_build_flow_scope *scope; + + scope = &lp_build_flow_peek(flow, lP_BUILD_FLOW_SCOPE)->scope; + if(!scope) + return; - if(!mask->skip_block) { - LLVMValueRef function = LLVMGetBasicBlockParent(current_block); - mask->skip_block = LLVMAppendBasicBlock(function, "skip"); + assert(*variable); + if(!*variable) + return; + + assert(flow->num_variables < LP_BUILD_FLOW_MAX_VARIABLES); + if(flow->num_variables >= LP_BUILD_FLOW_MAX_VARIABLES) + return; + + flow->variables[flow->num_variables++] = variable; + ++scope->num_variables; +} + + +void +lp_build_flow_scope_end(struct lp_build_flow_context *flow) +{ + struct lp_build_flow_scope *scope; + + scope = &lp_build_flow_pop(flow, lP_BUILD_FLOW_SCOPE)->scope; + if(!scope) + return; - mask->phi = LLVMBuildPhi(mask->builder, LLVMTypeOf(mask->value), ""); + assert(flow->num_variables >= scope->num_variables); + if(flow->num_variables < scope->num_variables) { + flow->num_variables = 0; + return; } + flow->num_variables -= scope->num_variables; +} + + +static LLVMBasicBlockRef +lp_build_flow_insert_block(struct lp_build_flow_context *flow) +{ + LLVMBasicBlockRef current_block; + LLVMBasicBlockRef next_block; + LLVMBasicBlockRef new_block; + + current_block = LLVMGetInsertBlock(flow->builder); + next_block = LLVMGetNextBasicBlock(current_block); - assert(next_block); if(next_block) { new_block = LLVMInsertBasicBlock(next_block, ""); } @@ -93,30 +295,148 @@ lp_build_mask_update(struct lp_build_mask_context *mask, new_block = LLVMAppendBasicBlock(function, ""); } - LLVMAddIncoming(mask->phi, &mask->value, ¤t_block, 1); - LLVMBuildCondBr(mask->builder, cond, mask->skip_block, new_block); + return new_block; +} + +void +lp_build_flow_skip_begin(struct lp_build_flow_context *flow) +{ + struct lp_build_flow_skip *skip; + LLVMBuilderRef builder; + unsigned i; + + skip = &lp_build_flow_push(flow, LP_BUILD_FLOW_SKIP)->skip; + if(!skip) + return; + + skip->block = lp_build_flow_insert_block(flow); + skip->num_variables = flow->num_variables; + if(!skip->num_variables) { + skip->phi = NULL; + return; + } - LLVMPositionBuilderAtEnd(mask->builder, new_block); -#endif + skip->phi = MALLOC(skip->num_variables * sizeof *skip->phi); + if(!skip->phi) { + skip->num_variables = 0; + return; + } + + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, skip->block); + + for(i = 0; i < skip->num_variables; ++i) + skip->phi[i] = LLVMBuildPhi(builder, LLVMTypeOf(*flow->variables[i]), ""); + + LLVMDisposeBuilder(builder); } -LLVMValueRef -lp_build_mask_end(struct lp_build_mask_context *mask) +void +lp_build_flow_skip_cond_break(struct lp_build_flow_context *flow, + LLVMValueRef cond) +{ + struct lp_build_flow_skip *skip; + LLVMBasicBlockRef current_block; + LLVMBasicBlockRef new_block; + unsigned i; + + skip = &lp_build_flow_peek(flow, LP_BUILD_FLOW_SKIP)->skip; + if(!skip) + return; + + current_block = LLVMGetInsertBlock(flow->builder); + + new_block = lp_build_flow_insert_block(flow); + + for(i = 0; i < skip->num_variables; ++i) { + assert(*flow->variables[i]); + LLVMAddIncoming(skip->phi[i], flow->variables[i], ¤t_block, 1); + } + + LLVMBuildCondBr(flow->builder, cond, skip->block, new_block); + + LLVMPositionBuilderAtEnd(flow->builder, new_block); + } + + +void +lp_build_flow_skip_end(struct lp_build_flow_context *flow) { - if(mask->skip_block) { - LLVMBasicBlockRef current_block = LLVMGetInsertBlock(mask->builder); + struct lp_build_flow_skip *skip; + LLVMBasicBlockRef current_block; + unsigned i; - LLVMAddIncoming(mask->phi, &mask->value, ¤t_block, 1); - LLVMBuildBr(mask->builder, mask->skip_block); + skip = &lp_build_flow_pop(flow, LP_BUILD_FLOW_SKIP)->skip; + if(!skip) + return; - LLVMPositionBuilderAtEnd(mask->builder, mask->skip_block); + current_block = LLVMGetInsertBlock(flow->builder); - mask->value = mask->phi; - mask->phi = NULL; - mask->skip_block = NULL; + for(i = 0; i < skip->num_variables; ++i) { + assert(*flow->variables[i]); + LLVMAddIncoming(skip->phi[i], flow->variables[i], ¤t_block, 1); + *flow->variables[i] = skip->phi[i]; } + LLVMBuildBr(flow->builder, skip->block); + LLVMPositionBuilderAtEnd(flow->builder, skip->block); + + FREE(skip->phi); +} + + +static void +lp_build_mask_check(struct lp_build_mask_context *mask) +{ + LLVMBuilderRef builder = mask->flow->builder; + LLVMValueRef cond; + + cond = LLVMBuildICmp(builder, + LLVMIntEQ, + LLVMBuildBitCast(builder, mask->value, mask->reg_type, ""), + LLVMConstNull(mask->reg_type), + ""); + + lp_build_flow_skip_cond_break(mask->flow, cond); +} + + +void +lp_build_mask_begin(struct lp_build_mask_context *mask, + struct lp_build_flow_context *flow, + struct lp_type type, + LLVMValueRef value) +{ + memset(mask, 0, sizeof *mask); + + mask->flow = flow; + mask->reg_type = LLVMIntType(type.width * type.length); + mask->value = value; + + lp_build_flow_scope_begin(flow); + lp_build_flow_scope_declare(flow, &mask->value); + lp_build_flow_skip_begin(flow); + + lp_build_mask_check(mask); +} + + +void +lp_build_mask_update(struct lp_build_mask_context *mask, + LLVMValueRef value) +{ + mask->value = LLVMBuildAnd( mask->flow->builder, mask->value, value, ""); + + lp_build_mask_check(mask); +} + + +LLVMValueRef +lp_build_mask_end(struct lp_build_mask_context *mask) +{ + lp_build_flow_skip_end(mask->flow); + lp_build_flow_scope_end(mask->flow); return mask->value; } diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.h b/src/gallium/drivers/llvmpipe/lp_bld_flow.h index 1b634ff038..e61999ff06 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_flow.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.h @@ -38,27 +38,53 @@ #include <llvm-c/Core.h> -union lp_type; +struct lp_type; + + +struct lp_build_flow_context; + + +struct lp_build_flow_context * +lp_build_flow_create(LLVMBuilderRef builder); + +void +lp_build_flow_destroy(struct lp_build_flow_context *flow); + +void +lp_build_flow_scope_begin(struct lp_build_flow_context *flow); + +void +lp_build_flow_scope_declare(struct lp_build_flow_context *flow, + LLVMValueRef *variable); + +void +lp_build_flow_scope_end(struct lp_build_flow_context *flow); + +void +lp_build_flow_skip_begin(struct lp_build_flow_context *flow); + +void +lp_build_flow_skip_cond_break(struct lp_build_flow_context *flow, + LLVMValueRef cond); + +void +lp_build_flow_skip_end(struct lp_build_flow_context *flow); struct lp_build_mask_context { - LLVMBuilderRef builder; + struct lp_build_flow_context *flow; LLVMTypeRef reg_type; LLVMValueRef value; - - LLVMValueRef phi; - - LLVMBasicBlockRef skip_block; }; void lp_build_mask_begin(struct lp_build_mask_context *mask, - LLVMBuilderRef builder, - union lp_type type, + struct lp_build_flow_context *flow, + struct lp_type type, LLVMValueRef value); /** diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format.h b/src/gallium/drivers/llvmpipe/lp_bld_format.h index 01c8a752d1..970bee379f 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_format.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_format.h @@ -25,77 +25,59 @@ * **************************************************************************/ -#ifndef LP_BLD_H -#define LP_BLD_H +#ifndef LP_BLD_FORMAT_H +#define LP_BLD_FORMAT_H /** * @file - * LLVM IR building helpers interfaces. - * - * We use LLVM-C bindings for now. They are not documented, but follow the C++ - * interfaces very closely, and appear to be complete enough for code - * genration. See - * http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html - * for a standalone example. + * Pixel format helpers. */ #include <llvm-c/Core.h> - + #include "pipe/p_format.h" +struct util_format_description; +struct lp_type; -union lp_type; + +boolean +lp_format_is_rgba8(const struct util_format_description *desc); + + +void +lp_build_format_swizzle_soa(const struct util_format_description *format_desc, + struct lp_type type, + const LLVMValueRef *unswizzled, + LLVMValueRef *swizzled); -/** - * Unpack a pixel into its RGBA components. - * - * @param packed integer. - * - * @return RGBA in a 4 floats vector. - */ LLVMValueRef -lp_build_unpack_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef packed); +lp_build_unpack_rgba_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + LLVMValueRef packed); -/** - * Pack a pixel. - * - * @param rgba 4 float vector with the unpacked components. - */ LLVMValueRef -lp_build_pack_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef rgba); +lp_build_unpack_rgba8_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + struct lp_type type, + LLVMValueRef packed); -/** - * Load a pixel into its RGBA components. - * - * @param ptr value with the pointer to the packed pixel. Pointer type is - * irrelevant. - * - * @return RGBA in a 4 floats vector. - */ LLVMValueRef -lp_build_load_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef ptr); +lp_build_pack_rgba_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + LLVMValueRef rgba); -/** - * Store a pixel. - * - * @param rgba 4 float vector with the unpacked components. - */ -void -lp_build_store_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef ptr, - LLVMValueRef rgba); +void +lp_build_unpack_rgba_soa(LLVMBuilderRef builder, + const struct util_format_description *format_desc, + struct lp_type type, + LLVMValueRef packed, + LLVMValueRef *rgba); -#endif /* !LP_BLD_H */ +#endif /* !LP_BLD_FORMAT_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c index dcbc0076c7..5836e0173f 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c @@ -25,18 +25,39 @@ * **************************************************************************/ +/** + * @file + * AoS pixel format manipulation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "util/u_cpu_detect.h" #include "util/u_format.h" +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_logic.h" +#include "lp_bld_swizzle.h" #include "lp_bld_format.h" +/** + * Unpack a single pixel into its RGBA components. + * + * @param packed integer. + * + * @return RGBA in a 4 floats vector. + * + * XXX: This is mostly for reference and testing -- operating a single pixel at + * a time is rarely if ever needed. + */ LLVMValueRef -lp_build_unpack_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef packed) +lp_build_unpack_rgba_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + LLVMValueRef packed) { - const struct util_format_description *desc; LLVMTypeRef type; LLVMValueRef shifted, casted, scaled, masked; LLVMValueRef shifts[4]; @@ -49,8 +70,6 @@ lp_build_unpack_rgba(LLVMBuilderRef builder, unsigned shift; unsigned i; - desc = util_format_description(format); - /* FIXME: Support more formats */ assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); assert(desc->block.width == 1); @@ -151,12 +170,130 @@ lp_build_unpack_rgba(LLVMBuilderRef builder, } +/** + * Take a vector with packed pixels and unpack into a rgba8 vector. + * + * Formats with bit depth smaller than 32bits are accepted, but they must be + * padded to 32bits. + */ LLVMValueRef -lp_build_pack_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef rgba) +lp_build_unpack_rgba8_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + struct lp_type type, + LLVMValueRef packed) +{ + struct lp_build_context bld; + bool rgba8; + LLVMValueRef res; + unsigned i; + + lp_build_context_init(&bld, builder, type); + + /* FIXME: Support more formats */ + assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(desc->block.width == 1); + assert(desc->block.height == 1); + assert(desc->block.bits <= 32); + + assert(!type.floating); + assert(!type.fixed); + assert(type.norm); + assert(type.width == 8); + assert(type.length % 4 == 0); + + rgba8 = TRUE; + for(i = 0; i < 4; ++i) { + assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED || + desc->channel[i].type == UTIL_FORMAT_TYPE_VOID); + if(desc->channel[0].size != 8) + rgba8 = FALSE; + } + + if(rgba8) { + /* + * The pixel is already in a rgba8 format variant. All it is necessary + * is to swizzle the channels. + */ + + unsigned char swizzles[4]; + boolean zeros[4]; /* bitwise AND mask */ + boolean ones[4]; /* bitwise OR mask */ + boolean swizzles_needed = FALSE; + boolean zeros_needed = FALSE; + boolean ones_needed = FALSE; + + for(i = 0; i < 4; ++i) { + enum util_format_swizzle swizzle = desc->swizzle[i]; + + /* Initialize with the no-op case */ + swizzles[i] = util_cpu_caps.little_endian ? 3 - i : i; + zeros[i] = TRUE; + ones[i] = FALSE; + + switch (swizzle) { + case UTIL_FORMAT_SWIZZLE_X: + case UTIL_FORMAT_SWIZZLE_Y: + case UTIL_FORMAT_SWIZZLE_Z: + case UTIL_FORMAT_SWIZZLE_W: + if(swizzle != swizzles[i]) { + swizzles[i] = swizzle; + swizzles_needed = TRUE; + } + break; + case UTIL_FORMAT_SWIZZLE_0: + zeros[i] = FALSE; + zeros_needed = TRUE; + break; + case UTIL_FORMAT_SWIZZLE_1: + ones[i] = TRUE; + ones_needed = TRUE; + break; + case UTIL_FORMAT_SWIZZLE_NONE: + assert(0); + break; + } + } + + res = packed; + + if(swizzles_needed) + res = lp_build_swizzle1_aos(&bld, res, swizzles); + + if(zeros_needed) { + /* Mask out zero channels */ + LLVMValueRef mask = lp_build_const_mask_aos(type, zeros); + res = LLVMBuildAnd(builder, res, mask, ""); + } + + if(ones_needed) { + /* Or one channels */ + LLVMValueRef mask = lp_build_const_mask_aos(type, ones); + res = LLVMBuildOr(builder, res, mask, ""); + } + } + else { + /* FIXME */ + assert(0); + res = lp_build_undef(type); + } + + return res; +} + + +/** + * Pack a single pixel. + * + * @param rgba 4 float vector with the unpacked components. + * + * XXX: This is mostly for reference and testing -- operating a single pixel at + * a time is rarely if ever needed. + */ +LLVMValueRef +lp_build_pack_rgba_aos(LLVMBuilderRef builder, + const struct util_format_description *desc, + LLVMValueRef rgba) { - const struct util_format_description *desc; LLVMTypeRef type; LLVMValueRef packed = NULL; LLVMValueRef swizzles[4]; @@ -167,8 +304,6 @@ lp_build_pack_rgba(LLVMBuilderRef builder, unsigned shift; unsigned i, j; - desc = util_format_description(format); - assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); assert(desc->block.width == 1); assert(desc->block.height == 1); @@ -247,57 +382,3 @@ lp_build_pack_rgba(LLVMBuilderRef builder, return packed; } - - -LLVMValueRef -lp_build_load_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef ptr) -{ - const struct util_format_description *desc; - LLVMTypeRef type; - LLVMValueRef packed; - - desc = util_format_description(format); - - /* FIXME: Support more formats */ - assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); - assert(desc->block.width == 1); - assert(desc->block.height == 1); - assert(desc->block.bits <= 32); - - type = LLVMIntType(desc->block.bits); - - ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, 0), ""); - - packed = LLVMBuildLoad(builder, ptr, ""); - - return lp_build_unpack_rgba(builder, format, packed); -} - - -void -lp_build_store_rgba(LLVMBuilderRef builder, - enum pipe_format format, - LLVMValueRef ptr, - LLVMValueRef rgba) -{ - const struct util_format_description *desc; - LLVMTypeRef type; - LLVMValueRef packed; - - desc = util_format_description(format); - - assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); - assert(desc->block.width == 1); - assert(desc->block.height == 1); - - type = LLVMIntType(desc->block.bits); - - packed = lp_build_pack_rgba(builder, format, rgba); - - ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, 0), ""); - - LLVMBuildStore(builder, packed, ptr); -} - diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_query.c b/src/gallium/drivers/llvmpipe/lp_bld_format_query.c new file mode 100644 index 0000000000..f3832d07ff --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_format_query.c @@ -0,0 +1,72 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Utility functions to make assertions about formats. + * + * This module centralizes most of logic used when determining what algorithm + * is most suitable (i.e., most efficient yet correct) for a given format. + * + * It might be possible to move some of these functions to u_format module, + * but since tiny differences in the format my render it more/less + * appropriate to a given algorithm it is impossible to make any long term + * guarantee about the semantics of these functions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_format.h" + +#include "lp_bld_format.h" + + +/** + * Whether this format is a 4 rgba8 variant + */ +boolean +lp_format_is_rgba8(const struct util_format_description *desc) +{ + unsigned chan; + + if(desc->block.width != 1 || + desc->block.height != 1 || + desc->block.bits != 32) + return FALSE; + + for(chan = 0; chan < 4; ++chan) { + if(desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED && + desc->channel[chan].type != UTIL_FORMAT_TYPE_SIGNED && + desc->channel[chan].type != UTIL_FORMAT_TYPE_VOID) + return FALSE; + if(desc->channel[chan].size != 8) + return FALSE; + } + + return TRUE; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c new file mode 100644 index 0000000000..64151d169d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c @@ -0,0 +1,149 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "util/u_format.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_conv.h" +#include "lp_bld_format.h" + + +static LLVMValueRef +lp_build_format_swizzle_chan_soa(struct lp_type type, + const LLVMValueRef *unswizzled, + enum util_format_swizzle swizzle) +{ + switch (swizzle) { + case UTIL_FORMAT_SWIZZLE_X: + case UTIL_FORMAT_SWIZZLE_Y: + case UTIL_FORMAT_SWIZZLE_Z: + case UTIL_FORMAT_SWIZZLE_W: + return unswizzled[swizzle]; + case UTIL_FORMAT_SWIZZLE_0: + return lp_build_zero(type); + case UTIL_FORMAT_SWIZZLE_1: + return lp_build_one(type); + case UTIL_FORMAT_SWIZZLE_NONE: + return lp_build_undef(type); + default: + assert(0); + return lp_build_undef(type); + } +} + + +void +lp_build_format_swizzle_soa(const struct util_format_description *format_desc, + struct lp_type type, + const LLVMValueRef *unswizzled, + LLVMValueRef *swizzled) +{ + if(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) { + enum util_format_swizzle swizzle = format_desc->swizzle[0]; + LLVMValueRef depth = lp_build_format_swizzle_chan_soa(type, unswizzled, swizzle); + swizzled[2] = swizzled[1] = swizzled[0] = depth; + swizzled[3] = lp_build_one(type); + } + else { + unsigned chan; + for (chan = 0; chan < 4; ++chan) { + enum util_format_swizzle swizzle = format_desc->swizzle[chan]; + swizzled[chan] = lp_build_format_swizzle_chan_soa(type, unswizzled, swizzle); + } + } +} + + +void +lp_build_unpack_rgba_soa(LLVMBuilderRef builder, + const struct util_format_description *format_desc, + struct lp_type type, + LLVMValueRef packed, + LLVMValueRef *rgba) +{ + LLVMValueRef inputs[4]; + unsigned start; + unsigned chan; + + /* FIXME: Support more formats */ + assert(format_desc->layout == UTIL_FORMAT_LAYOUT_ARITH || + (format_desc->layout == UTIL_FORMAT_LAYOUT_ARRAY && + format_desc->block.bits == format_desc->channel[0].size)); + assert(format_desc->block.width == 1); + assert(format_desc->block.height == 1); + assert(format_desc->block.bits <= 32); + + /* Decode the input vector components */ + start = 0; + for (chan = 0; chan < 4; ++chan) { + unsigned width = format_desc->channel[chan].size; + unsigned stop = start + width; + LLVMValueRef input; + + input = packed; + + switch(format_desc->channel[chan].type) { + case UTIL_FORMAT_TYPE_VOID: + input = NULL; + break; + + case UTIL_FORMAT_TYPE_UNSIGNED: + if(type.floating) { + if(start) + input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(type, start), ""); + if(stop < format_desc->block.bits) { + unsigned mask = ((unsigned long long)1 << width) - 1; + input = LLVMBuildAnd(builder, input, lp_build_int_const_scalar(type, mask), ""); + } + + if(format_desc->channel[chan].normalized) + input = lp_build_unsigned_norm_to_float(builder, width, type, input); + else + input = LLVMBuildFPToSI(builder, input, lp_build_vec_type(type), ""); + } + else { + /* FIXME */ + assert(0); + input = lp_build_undef(type); + } + break; + + default: + /* fall through */ + input = lp_build_undef(type); + break; + } + + inputs[chan] = input; + + start = stop; + } + + lp_build_format_swizzle_soa(format_desc, type, inputs, rgba); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c index cfe20a0d75..818c0e943e 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_interp.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c @@ -109,32 +109,6 @@ coeffs_init(struct lp_build_interp_soa_context *bld, /** - * Small vector x scale multiplication optimization. - * - * TODO: Should be elsewhere. - */ -static LLVMValueRef -coeff_multiply(struct lp_build_interp_soa_context *bld, - LLVMValueRef coeff, - int step) -{ - LLVMValueRef factor; - - switch(step) { - case 0: - return bld->base.zero; - case 1: - return coeff; - case 2: - return lp_build_add(&bld->base, coeff, coeff); - default: - factor = lp_build_const_scalar(bld->base.type, (double)step); - return lp_build_mul(&bld->base, coeff, factor); - } -} - - -/** * Multiply the dadx and dady with the xstep and ystep respectively. */ static void @@ -149,8 +123,8 @@ coeffs_update(struct lp_build_interp_soa_context *bld) if (mode != TGSI_INTERPOLATE_CONSTANT) { for(chan = 0; chan < NUM_CHANNELS; ++chan) { if(mask & (1 << chan)) { - bld->dadx[attrib][chan] = coeff_multiply(bld, bld->dadx[attrib][chan], bld->xstep); - bld->dady[attrib][chan] = coeff_multiply(bld, bld->dady[attrib][chan], bld->ystep); + bld->dadx[attrib][chan] = lp_build_mul_imm(&bld->base, bld->dadx[attrib][chan], bld->xstep); + bld->dady[attrib][chan] = lp_build_mul_imm(&bld->base, bld->dady[attrib][chan], bld->ystep); } } } @@ -292,7 +266,7 @@ void lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld, const struct tgsi_token *tokens, LLVMBuilderRef builder, - union lp_type type, + struct lp_type type, LLVMValueRef a0_ptr, LLVMValueRef dadx_ptr, LLVMValueRef dady_ptr, diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.h b/src/gallium/drivers/llvmpipe/lp_bld_interp.h index 9194f6233a..9c57a10879 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_interp.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.h @@ -83,7 +83,7 @@ void lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld, const struct tgsi_token *tokens, LLVMBuilderRef builder, - union lp_type type, + struct lp_type type, LLVMValueRef a0_ptr, LLVMValueRef dadx_ptr, LLVMValueRef dady_ptr, diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.c b/src/gallium/drivers/llvmpipe/lp_bld_logic.c index 8631efd6c3..db22a8028a 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_logic.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.c @@ -33,6 +33,8 @@ */ +#include "util/u_cpu_detect.h" + #include "lp_bld_type.h" #include "lp_bld_const.h" #include "lp_bld_intr.h" @@ -45,7 +47,7 @@ lp_build_cmp(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); LLVMValueRef zeros = LLVMConstNull(int_vec_type); @@ -65,7 +67,7 @@ lp_build_cmp(struct lp_build_context *bld, #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) if(type.width * type.length == 128) { - if(type.floating) { + if(type.floating && util_cpu_caps.has_sse) { LLVMValueRef args[3]; unsigned cc; boolean swap; @@ -114,7 +116,7 @@ lp_build_cmp(struct lp_build_context *bld, res = LLVMBuildBitCast(bld->builder, res, int_vec_type, ""); return res; } - else { + else if(util_cpu_caps.has_sse2) { static const struct { unsigned swap:1; unsigned eq:1; @@ -301,7 +303,7 @@ lp_build_select(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { - union lp_type type = bld->type; + struct lp_type type = bld->type; LLVMValueRef res; if(a == b) @@ -313,8 +315,6 @@ lp_build_select(struct lp_build_context *bld, b = LLVMBuildBitCast(bld->builder, b, int_vec_type, ""); } - /* TODO: On SSE4 we could do this with a single instruction -- PBLENDVB */ - a = LLVMBuildAnd(bld->builder, a, mask, ""); /* This often gets translated to PANDN, but sometimes the NOT is @@ -339,9 +339,9 @@ LLVMValueRef lp_build_select_aos(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b, - boolean cond[4]) + const boolean cond[4]) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; const unsigned n = type.length; unsigned i, j; @@ -376,9 +376,9 @@ lp_build_select_aos(struct lp_build_context *bld, return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), ""); } + else { #if 0 - else if(0) { - /* FIXME: Unfortunately select of vectors do not work */ + /* XXX: Unfortunately select of vectors do not work */ /* Use a select */ LLVMTypeRef elem_type = LLVMInt1Type(); LLVMValueRef cond[LP_MAX_VECTOR_LENGTH]; @@ -388,10 +388,9 @@ lp_build_select_aos(struct lp_build_context *bld, cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0); return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, ""); - } -#endif - else { +#else LLVMValueRef mask = lp_build_const_mask_aos(type, cond); return lp_build_select(bld, mask, a, b); +#endif } } diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.h b/src/gallium/drivers/llvmpipe/lp_bld_logic.h index 29b9e1c45b..d67500ef70 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_logic.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.h @@ -42,7 +42,7 @@ #include "pipe/p_defines.h" /* For PIPE_FUNC_xxx */ -union lp_type type; +struct lp_type; struct lp_build_context; @@ -66,7 +66,7 @@ LLVMValueRef lp_build_select_aos(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b, - boolean cond[4]); + const boolean cond[4]); #endif /* !LP_BLD_LOGIC_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_misc.cpp b/src/gallium/drivers/llvmpipe/lp_bld_misc.cpp new file mode 100644 index 0000000000..d3f78c06d9 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_misc.cpp @@ -0,0 +1,61 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "pipe/p_config.h" + +#include "lp_bld_misc.h" + + +#ifndef LLVM_NATIVE_ARCH + +namespace llvm { + extern void LinkInJIT(); +} + + +void +LLVMLinkInJIT(void) +{ + llvm::LinkInJIT(); +} + + +extern "C" int X86TargetMachineModule; + + +int +LLVMInitializeNativeTarget(void) +{ +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + X86TargetMachineModule = 1; +#endif + return 0; +} + + +#endif diff --git a/src/gallium/drivers/llvmpipe/lp_bld_misc.h b/src/gallium/drivers/llvmpipe/lp_bld_misc.h new file mode 100644 index 0000000000..0e787e0b9c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_misc.h @@ -0,0 +1,56 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef LP_BLD_MISC_H +#define LP_BLD_MISC_H + + +#include "llvm/Config/config.h" + +#ifdef __cplusplus +extern "C" { +#endif + + +#ifndef LLVM_NATIVE_ARCH + +void +LLVMLinkInJIT(void); + +int +LLVMInitializeNativeTarget(void); + +#endif /* !LLVM_NATIVE_ARCH */ + + +#ifdef __cplusplus +} +#endif + + +#endif /* !LP_BLD_MISC_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_pack.c b/src/gallium/drivers/llvmpipe/lp_bld_pack.c new file mode 100644 index 0000000000..bc360ad77a --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_pack.c @@ -0,0 +1,418 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helper functions for packing/unpacking. + * + * Pack/unpacking is necessary for conversion between types of different + * bit width. + * + * They are also commonly used when an computation needs higher + * precision for the intermediate values. For example, if one needs the + * function: + * + * c = compute(a, b); + * + * to use more precision for intermediate results then one should implement it + * as: + * + * LLVMValueRef + * compute(LLVMBuilderRef builder struct lp_type type, LLVMValueRef a, LLVMValueRef b) + * { + * struct lp_type wide_type = lp_wider_type(type); + * LLVMValueRef al, ah, bl, bh, cl, ch, c; + * + * lp_build_unpack2(builder, type, wide_type, a, &al, &ah); + * lp_build_unpack2(builder, type, wide_type, b, &bl, &bh); + * + * cl = compute_half(al, bl); + * ch = compute_half(ah, bh); + * + * c = lp_build_pack2(bld->builder, wide_type, type, cl, ch); + * + * return c; + * } + * + * where compute_half() would do the computation for half the elements with + * twice the precision. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_debug.h" +#include "util/u_math.h" +#include "util/u_cpu_detect.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_intr.h" +#include "lp_bld_arit.h" +#include "lp_bld_pack.h" + + +/** + * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions. + */ +static LLVMValueRef +lp_build_const_unpack_shuffle(unsigned n, unsigned lo_hi) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i, j; + + assert(n <= LP_MAX_VECTOR_LENGTH); + assert(lo_hi < 2); + + /* TODO: cache results in a static table */ + + for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) { + elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0); + elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0); + } + + return LLVMConstVector(elems, n); +} + + +/** + * Build shuffle vectors that match PACKxx instructions. + */ +static LLVMValueRef +lp_build_const_pack_shuffle(unsigned n) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(n <= LP_MAX_VECTOR_LENGTH); + + /* TODO: cache results in a static table */ + + for(i = 0; i < n; ++i) + elems[i] = LLVMConstInt(LLVMInt32Type(), 2*i, 0); + + return LLVMConstVector(elems, n); +} + + +/** + * Interleave vector elements. + * + * Matches the PUNPCKLxx and PUNPCKHxx SSE instructions. + */ +LLVMValueRef +lp_build_interleave2(LLVMBuilderRef builder, + struct lp_type type, + LLVMValueRef a, + LLVMValueRef b, + unsigned lo_hi) +{ + LLVMValueRef shuffle; + + shuffle = lp_build_const_unpack_shuffle(type.length, lo_hi); + + return LLVMBuildShuffleVector(builder, a, b, shuffle, ""); +} + + +/** + * Double the bit width. + * + * This will only change the number of bits the values are represented, not the + * values themselves. + */ +void +lp_build_unpack2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef src, + LLVMValueRef *dst_lo, + LLVMValueRef *dst_hi) +{ + LLVMValueRef msb; + LLVMTypeRef dst_vec_type; + + assert(!src_type.floating); + assert(!dst_type.floating); + assert(dst_type.width == src_type.width * 2); + assert(dst_type.length * 2 == src_type.length); + + if(dst_type.sign && src_type.sign) { + /* Replicate the sign bit in the most significant bits */ + msb = LLVMBuildAShr(builder, src, lp_build_int_const_scalar(src_type, src_type.width - 1), ""); + } + else + /* Most significant bits always zero */ + msb = lp_build_zero(src_type); + + /* Interleave bits */ + if(util_cpu_caps.little_endian) { + *dst_lo = lp_build_interleave2(builder, src_type, src, msb, 0); + *dst_hi = lp_build_interleave2(builder, src_type, src, msb, 1); + } + else { + *dst_lo = lp_build_interleave2(builder, src_type, msb, src, 0); + *dst_hi = lp_build_interleave2(builder, src_type, msb, src, 1); + } + + /* Cast the result into the new type (twice as wide) */ + + dst_vec_type = lp_build_vec_type(dst_type); + + *dst_lo = LLVMBuildBitCast(builder, *dst_lo, dst_vec_type, ""); + *dst_hi = LLVMBuildBitCast(builder, *dst_hi, dst_vec_type, ""); +} + + +/** + * Expand the bit width. + * + * This will only change the number of bits the values are represented, not the + * values themselves. + */ +void +lp_build_unpack(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef src, + LLVMValueRef *dst, unsigned num_dsts) +{ + unsigned num_tmps; + unsigned i; + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length == dst_type.length * num_dsts); + + num_tmps = 1; + dst[0] = src; + + while(src_type.width < dst_type.width) { + struct lp_type tmp_type = src_type; + + tmp_type.width *= 2; + tmp_type.length /= 2; + + for(i = num_tmps; i--; ) { + lp_build_unpack2(builder, src_type, tmp_type, dst[i], &dst[2*i + 0], &dst[2*i + 1]); + } + + src_type = tmp_type; + + num_tmps *= 2; + } + + assert(num_tmps == num_dsts); +} + + +/** + * Non-interleaved pack. + * + * This will move values as + * + * lo = __ l0 __ l1 __ l2 __.. __ ln + * hi = __ h0 __ h1 __ h2 __.. __ hn + * res = l0 l1 l2 .. ln h0 h1 h2 .. hn + * + * This will only change the number of bits the values are represented, not the + * values themselves. + * + * It is assumed the values are already clamped into the destination type range. + * Values outside that range will produce undefined results. Use + * lp_build_packs2 instead. + */ +LLVMValueRef +lp_build_pack2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef lo, + LLVMValueRef hi) +{ + LLVMTypeRef src_vec_type = lp_build_vec_type(src_type); + LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type); + LLVMValueRef shuffle; + LLVMValueRef res; + + dst_vec_type = lp_build_vec_type(dst_type); + + assert(!src_type.floating); + assert(!dst_type.floating); + assert(src_type.width == dst_type.width * 2); + assert(src_type.length * 2 == dst_type.length); + + if(util_cpu_caps.has_sse2 && src_type.width * src_type.length == 128) { + switch(src_type.width) { + case 32: + if(dst_type.sign) { + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", src_vec_type, lo, hi); + } + else { + if (util_cpu_caps.has_sse4_1) { + /* PACKUSDW is the only instrinsic with a consistent signature */ + return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi); + } + else { + assert(0); + return LLVMGetUndef(dst_vec_type); + } + } + break; + + case 16: + if(dst_type.sign) + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", src_vec_type, lo, hi); + else + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", src_vec_type, lo, hi); + break; + + default: + assert(0); + return LLVMGetUndef(dst_vec_type); + break; + } + + res = LLVMBuildBitCast(builder, res, dst_vec_type, ""); + return res; + } + + lo = LLVMBuildBitCast(builder, lo, dst_vec_type, ""); + hi = LLVMBuildBitCast(builder, hi, dst_vec_type, ""); + + shuffle = lp_build_const_pack_shuffle(dst_type.length); + + res = LLVMBuildShuffleVector(builder, lo, hi, shuffle, ""); + + return res; +} + + + +/** + * Non-interleaved pack and saturate. + * + * Same as lp_build_pack2 but will saturate values so that they fit into the + * destination type. + */ +LLVMValueRef +lp_build_packs2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef lo, + LLVMValueRef hi) +{ + boolean clamp; + + assert(!src_type.floating); + assert(!dst_type.floating); + assert(src_type.sign == dst_type.sign); + assert(src_type.width == dst_type.width * 2); + assert(src_type.length * 2 == dst_type.length); + + clamp = TRUE; + + /* All X86 SSE non-interleaved pack instructions take signed inputs and + * saturate them, so no need to clamp for those cases. */ + if(util_cpu_caps.has_sse2 && + src_type.width * src_type.length == 128 && + src_type.sign) + clamp = FALSE; + + if(clamp) { + struct lp_build_context bld; + unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width; + LLVMValueRef dst_max = lp_build_int_const_scalar(src_type, ((unsigned long long)1 << dst_bits) - 1); + lp_build_context_init(&bld, builder, src_type); + lo = lp_build_min(&bld, lo, dst_max); + hi = lp_build_min(&bld, hi, dst_max); + /* FIXME: What about lower bound? */ + } + + return lp_build_pack2(builder, src_type, dst_type, lo, hi); +} + + +/** + * Truncate the bit width. + * + * TODO: Handle saturation consistently. + */ +LLVMValueRef +lp_build_pack(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + boolean clamped, + const LLVMValueRef *src, unsigned num_srcs) +{ + LLVMValueRef (*pack2)(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef lo, + LLVMValueRef hi); + LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * num_srcs == dst_type.length); + + if(clamped) + pack2 = &lp_build_pack2; + else + pack2 = &lp_build_packs2; + + for(i = 0; i < num_srcs; ++i) + tmp[i] = src[i]; + + while(src_type.width > dst_type.width) { + struct lp_type tmp_type = src_type; + + tmp_type.width /= 2; + tmp_type.length *= 2; + + /* Take in consideration the sign changes only in the last step */ + if(tmp_type.width == dst_type.width) + tmp_type.sign = dst_type.sign; + + num_srcs /= 2; + + for(i = 0; i < num_srcs; ++i) + tmp[i] = pack2(builder, src_type, tmp_type, tmp[2*i + 0], tmp[2*i + 1]); + + src_type = tmp_type; + } + + assert(num_srcs == 1); + + return tmp[0]; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_pack.h b/src/gallium/drivers/llvmpipe/lp_bld_pack.h new file mode 100644 index 0000000000..fb2a34984a --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_pack.h @@ -0,0 +1,95 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for packing/unpacking conversions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_PACK_H +#define LP_BLD_PACK_H + + +#include <llvm-c/Core.h> + + +struct lp_type; + + +LLVMValueRef +lp_build_interleave2(LLVMBuilderRef builder, + struct lp_type type, + LLVMValueRef a, + LLVMValueRef b, + unsigned lo_hi); + + +void +lp_build_unpack2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef src, + LLVMValueRef *dst_lo, + LLVMValueRef *dst_hi); + + +void +lp_build_unpack(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef src, + LLVMValueRef *dst, unsigned num_dsts); + + +LLVMValueRef +lp_build_packs2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef lo, + LLVMValueRef hi); + + +LLVMValueRef +lp_build_pack2(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + LLVMValueRef lo, + LLVMValueRef hi); + + +LLVMValueRef +lp_build_pack(LLVMBuilderRef builder, + struct lp_type src_type, + struct lp_type dst_type, + boolean clamped, + const LLVMValueRef *src, unsigned num_srcs); + + +#endif /* !LP_BLD_PACK_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample.c b/src/gallium/drivers/llvmpipe/lp_bld_sample.c new file mode 100644 index 0000000000..4d272bea87 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_sample.c @@ -0,0 +1,190 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Texture sampling -- common code. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_defines.h" +#include "pipe/p_state.h" +#include "util/u_format.h" +#include "util/u_math.h" +#include "lp_bld_debug.h" +#include "lp_bld_const.h" +#include "lp_bld_arit.h" +#include "lp_bld_type.h" +#include "lp_bld_format.h" +#include "lp_bld_sample.h" + + +void +lp_sampler_static_state(struct lp_sampler_static_state *state, + const struct pipe_texture *texture, + const struct pipe_sampler_state *sampler) +{ + memset(state, 0, sizeof *state); + + if(!texture) + return; + + if(!sampler) + return; + + state->format = texture->format; + state->target = texture->target; + state->pot_width = util_is_pot(texture->width[0]); + state->pot_height = util_is_pot(texture->height[0]); + state->pot_depth = util_is_pot(texture->depth[0]); + + state->wrap_s = sampler->wrap_s; + state->wrap_t = sampler->wrap_t; + state->wrap_r = sampler->wrap_r; + state->min_img_filter = sampler->min_img_filter; + state->min_mip_filter = sampler->min_mip_filter; + state->mag_img_filter = sampler->mag_img_filter; + if(sampler->compare_mode) { + state->compare_mode = sampler->compare_mode; + state->compare_func = sampler->compare_func; + } + state->normalized_coords = sampler->normalized_coords; + state->prefilter = sampler->prefilter; +} + + +/** + * Gather elements from scatter positions in memory into a single vector. + * + * @param src_width src element width + * @param dst_width result element width (source will be expanded to fit) + * @param length length of the offsets, + * @param base_ptr base pointer, should be a i8 pointer type. + * @param offsets vector with offsets + */ +LLVMValueRef +lp_build_gather(LLVMBuilderRef builder, + unsigned length, + unsigned src_width, + unsigned dst_width, + LLVMValueRef base_ptr, + LLVMValueRef offsets) +{ + LLVMTypeRef src_type = LLVMIntType(src_width); + LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0); + LLVMTypeRef dst_elem_type = LLVMIntType(dst_width); + LLVMTypeRef dst_vec_type = LLVMVectorType(dst_elem_type, length); + LLVMValueRef res; + unsigned i; + + res = LLVMGetUndef(dst_vec_type); + for(i = 0; i < length; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef elem_offset; + LLVMValueRef elem_ptr; + LLVMValueRef elem; + + elem_offset = LLVMBuildExtractElement(builder, offsets, index, ""); + elem_ptr = LLVMBuildGEP(builder, base_ptr, &elem_offset, 1, ""); + elem_ptr = LLVMBuildBitCast(builder, elem_ptr, src_ptr_type, ""); + elem = LLVMBuildLoad(builder, elem_ptr, ""); + + assert(src_width <= dst_width); + if(src_width > dst_width) + elem = LLVMBuildTrunc(builder, elem, dst_elem_type, ""); + if(src_width < dst_width) + elem = LLVMBuildZExt(builder, elem, dst_elem_type, ""); + + res = LLVMBuildInsertElement(builder, res, elem, index, ""); + } + + return res; +} + + +/** + * Compute the offset of a pixel. + * + * x, y, y_stride are vectors + */ +LLVMValueRef +lp_build_sample_offset(struct lp_build_context *bld, + const struct util_format_description *format_desc, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr) +{ + LLVMValueRef x_stride; + LLVMValueRef offset; + + x_stride = lp_build_const_scalar(bld->type, format_desc->block.bits/8); + + if(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) { + LLVMValueRef x_lo, x_hi; + LLVMValueRef y_lo, y_hi; + LLVMValueRef x_stride_lo, x_stride_hi; + LLVMValueRef y_stride_lo, y_stride_hi; + LLVMValueRef x_offset_lo, x_offset_hi; + LLVMValueRef y_offset_lo, y_offset_hi; + LLVMValueRef offset_lo, offset_hi; + + x_lo = LLVMBuildAnd(bld->builder, x, bld->one, ""); + y_lo = LLVMBuildAnd(bld->builder, y, bld->one, ""); + + x_hi = LLVMBuildLShr(bld->builder, x, bld->one, ""); + y_hi = LLVMBuildLShr(bld->builder, y, bld->one, ""); + + x_stride_lo = x_stride; + y_stride_lo = lp_build_const_scalar(bld->type, 2*format_desc->block.bits/8); + + x_stride_hi = lp_build_const_scalar(bld->type, 4*format_desc->block.bits/8); + y_stride_hi = LLVMBuildShl(bld->builder, y_stride, bld->one, ""); + + x_offset_lo = lp_build_mul(bld, x_lo, x_stride_lo); + y_offset_lo = lp_build_mul(bld, y_lo, y_stride_lo); + offset_lo = lp_build_add(bld, x_offset_lo, y_offset_lo); + + x_offset_hi = lp_build_mul(bld, x_hi, x_stride_hi); + y_offset_hi = lp_build_mul(bld, y_hi, y_stride_hi); + offset_hi = lp_build_add(bld, x_offset_hi, y_offset_hi); + + offset = lp_build_add(bld, offset_hi, offset_lo); + } + else { + LLVMValueRef x_offset; + LLVMValueRef y_offset; + + x_offset = lp_build_mul(bld, x, x_stride); + y_offset = lp_build_mul(bld, y, y_stride); + + offset = lp_build_add(bld, x_offset, y_offset); + } + + return offset; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample.h b/src/gallium/drivers/llvmpipe/lp_bld_sample.h new file mode 100644 index 0000000000..8cb8210ca7 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_sample.h @@ -0,0 +1,155 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Texture sampling. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_SAMPLE_H +#define LP_BLD_SAMPLE_H + + +#include <llvm-c/Core.h> + +struct pipe_texture; +struct pipe_sampler_state; +struct util_format_description; +struct lp_type; +struct lp_build_context; + + +/** + * Sampler static state. + * + * These are the bits of state from pipe_texture and pipe_sampler_state that + * are embedded in the generated code. + */ +struct lp_sampler_static_state +{ + /* pipe_texture's state */ + enum pipe_format format; + unsigned target:2; + unsigned pot_width:1; + unsigned pot_height:1; + unsigned pot_depth:1; + + /* pipe_sampler_state's state */ + unsigned wrap_s:3; + unsigned wrap_t:3; + unsigned wrap_r:3; + unsigned min_img_filter:2; + unsigned min_mip_filter:2; + unsigned mag_img_filter:2; + unsigned compare_mode:1; + unsigned compare_func:3; + unsigned normalized_coords:1; + unsigned prefilter:4; +}; + + +/** + * Sampler dynamic state. + * + * These are the bits of state from pipe_texture and pipe_sampler_state that + * are computed in runtime. + * + * There are obtained through callbacks, as we don't want to tie the texture + * sampling code generation logic to any particular texture layout or pipe + * driver. + */ +struct lp_sampler_dynamic_state +{ + + /** Obtain the base texture width. */ + LLVMValueRef + (*width)( struct lp_sampler_dynamic_state *state, + LLVMBuilderRef builder, + unsigned unit); + + /** Obtain the base texture height. */ + LLVMValueRef + (*height)( struct lp_sampler_dynamic_state *state, + LLVMBuilderRef builder, + unsigned unit); + + LLVMValueRef + (*stride)( struct lp_sampler_dynamic_state *state, + LLVMBuilderRef builder, + unsigned unit); + + LLVMValueRef + (*data_ptr)( struct lp_sampler_dynamic_state *state, + LLVMBuilderRef builder, + unsigned unit); + +}; + + +/** + * Derive the sampler static state. + */ +void +lp_sampler_static_state(struct lp_sampler_static_state *state, + const struct pipe_texture *texture, + const struct pipe_sampler_state *sampler); + + +LLVMValueRef +lp_build_gather(LLVMBuilderRef builder, + unsigned length, + unsigned src_width, + unsigned dst_width, + LLVMValueRef base_ptr, + LLVMValueRef offsets); + + +LLVMValueRef +lp_build_sample_offset(struct lp_build_context *bld, + const struct util_format_description *format_desc, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr); + + +void +lp_build_sample_soa(LLVMBuilderRef builder, + const struct lp_sampler_static_state *static_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct lp_type fp_type, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel); + + + +#endif /* LP_BLD_SAMPLE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c new file mode 100644 index 0000000000..47b68b71e2 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c @@ -0,0 +1,595 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Texture sampling -- SoA. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_defines.h" +#include "pipe/p_state.h" +#include "util/u_debug.h" +#include "util/u_debug_dump.h" +#include "util/u_memory.h" +#include "util/u_math.h" +#include "util/u_format.h" +#include "util/u_cpu_detect.h" +#include "lp_bld_debug.h" +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_conv.h" +#include "lp_bld_arit.h" +#include "lp_bld_logic.h" +#include "lp_bld_swizzle.h" +#include "lp_bld_pack.h" +#include "lp_bld_format.h" +#include "lp_bld_sample.h" + + +/** + * Keep all information for sampling code generation in a single place. + */ +struct lp_build_sample_context +{ + LLVMBuilderRef builder; + + const struct lp_sampler_static_state *static_state; + + struct lp_sampler_dynamic_state *dynamic_state; + + const struct util_format_description *format_desc; + + /** Incoming coordinates type and build context */ + struct lp_type coord_type; + struct lp_build_context coord_bld; + + /** Integer coordinates */ + struct lp_type int_coord_type; + struct lp_build_context int_coord_bld; + + /** Output texels type and build context */ + struct lp_type texel_type; + struct lp_build_context texel_bld; +}; + + +static void +lp_build_sample_texel_soa(struct lp_build_sample_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr, + LLVMValueRef *texel) +{ + LLVMValueRef offset; + LLVMValueRef packed; + + offset = lp_build_sample_offset(&bld->int_coord_bld, + bld->format_desc, + x, y, y_stride, + data_ptr); + + assert(bld->format_desc->block.width == 1); + assert(bld->format_desc->block.height == 1); + assert(bld->format_desc->block.bits <= bld->texel_type.width); + + packed = lp_build_gather(bld->builder, + bld->texel_type.length, + bld->format_desc->block.bits, + bld->texel_type.width, + data_ptr, offset); + + lp_build_unpack_rgba_soa(bld->builder, + bld->format_desc, + bld->texel_type, + packed, texel); +} + + +static LLVMValueRef +lp_build_sample_packed(struct lp_build_sample_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr) +{ + LLVMValueRef offset; + + offset = lp_build_sample_offset(&bld->int_coord_bld, + bld->format_desc, + x, y, y_stride, + data_ptr); + + assert(bld->format_desc->block.width == 1); + assert(bld->format_desc->block.height == 1); + assert(bld->format_desc->block.bits <= bld->texel_type.width); + + return lp_build_gather(bld->builder, + bld->texel_type.length, + bld->format_desc->block.bits, + bld->texel_type.width, + data_ptr, offset); +} + + +static LLVMValueRef +lp_build_sample_wrap(struct lp_build_sample_context *bld, + LLVMValueRef coord, + LLVMValueRef length, + boolean is_pot, + unsigned wrap_mode) +{ + struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + LLVMValueRef length_minus_one; + + length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one); + + switch(wrap_mode) { + case PIPE_TEX_WRAP_REPEAT: + if(is_pot) + coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, ""); + else + /* Signed remainder won't give the right results for negative + * dividends but unsigned remainder does.*/ + coord = LLVMBuildURem(bld->builder, coord, length, ""); + break; + + case PIPE_TEX_WRAP_CLAMP: + coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero); + coord = lp_build_min(int_coord_bld, coord, length_minus_one); + break; + + case PIPE_TEX_WRAP_CLAMP_TO_EDGE: + case PIPE_TEX_WRAP_CLAMP_TO_BORDER: + case PIPE_TEX_WRAP_MIRROR_REPEAT: + case PIPE_TEX_WRAP_MIRROR_CLAMP: + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: + /* FIXME */ + _debug_printf("warning: failed to translate texture wrap mode %s\n", + debug_dump_tex_wrap(wrap_mode, TRUE)); + coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero); + coord = lp_build_min(int_coord_bld, coord, length_minus_one); + break; + + default: + assert(0); + } + + return coord; +} + + +static void +lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef width, + LLVMValueRef height, + LLVMValueRef stride, + LLVMValueRef data_ptr, + LLVMValueRef *texel) +{ + LLVMValueRef x; + LLVMValueRef y; + + x = lp_build_ifloor(&bld->coord_bld, s); + y = lp_build_ifloor(&bld->coord_bld, t); + + x = lp_build_sample_wrap(bld, x, width, bld->static_state->pot_width, bld->static_state->wrap_s); + y = lp_build_sample_wrap(bld, y, height, bld->static_state->pot_height, bld->static_state->wrap_t); + + lp_build_sample_texel_soa(bld, x, y, stride, data_ptr, texel); +} + + +static void +lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef width, + LLVMValueRef height, + LLVMValueRef stride, + LLVMValueRef data_ptr, + LLVMValueRef *texel) +{ + LLVMValueRef half; + LLVMValueRef s_ipart; + LLVMValueRef t_ipart; + LLVMValueRef s_fpart; + LLVMValueRef t_fpart; + LLVMValueRef x0, x1; + LLVMValueRef y0, y1; + LLVMValueRef neighbors[2][2][4]; + unsigned chan; + + half = lp_build_const_scalar(bld->coord_type, 0.5); + s = lp_build_sub(&bld->coord_bld, s, half); + t = lp_build_sub(&bld->coord_bld, t, half); + + s_ipart = lp_build_floor(&bld->coord_bld, s); + t_ipart = lp_build_floor(&bld->coord_bld, t); + + s_fpart = lp_build_sub(&bld->coord_bld, s, s_ipart); + t_fpart = lp_build_sub(&bld->coord_bld, t, t_ipart); + + x0 = lp_build_itrunc(&bld->coord_bld, s_ipart); + y0 = lp_build_itrunc(&bld->coord_bld, t_ipart); + + x0 = lp_build_sample_wrap(bld, x0, width, bld->static_state->pot_width, bld->static_state->wrap_s); + y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t); + + x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one); + y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one); + + x1 = lp_build_sample_wrap(bld, x1, width, bld->static_state->pot_width, bld->static_state->wrap_s); + y1 = lp_build_sample_wrap(bld, y1, height, bld->static_state->pot_height, bld->static_state->wrap_t); + + lp_build_sample_texel_soa(bld, x0, y0, stride, data_ptr, neighbors[0][0]); + lp_build_sample_texel_soa(bld, x1, y0, stride, data_ptr, neighbors[0][1]); + lp_build_sample_texel_soa(bld, x0, y1, stride, data_ptr, neighbors[1][0]); + lp_build_sample_texel_soa(bld, x1, y1, stride, data_ptr, neighbors[1][1]); + + /* TODO: Don't interpolate missing channels */ + for(chan = 0; chan < 4; ++chan) { + texel[chan] = lp_build_lerp_2d(&bld->texel_bld, + s_fpart, t_fpart, + neighbors[0][0][chan], + neighbors[0][1][chan], + neighbors[1][0][chan], + neighbors[1][1][chan]); + } +} + + +static void +lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder, + struct lp_type dst_type, + LLVMValueRef packed, + LLVMValueRef *rgba) +{ + LLVMValueRef mask = lp_build_int_const_scalar(dst_type, 0xff); + unsigned chan; + + /* Decode the input vector components */ + for (chan = 0; chan < 4; ++chan) { + unsigned start = chan*8; + unsigned stop = start + 8; + LLVMValueRef input; + + input = packed; + + if(start) + input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(dst_type, start), ""); + + if(stop < 32) + input = LLVMBuildAnd(builder, input, mask, ""); + + input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input); + + rgba[chan] = input; + } +} + + +static void +lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef width, + LLVMValueRef height, + LLVMValueRef stride, + LLVMValueRef data_ptr, + LLVMValueRef *texel) +{ + LLVMBuilderRef builder = bld->builder; + struct lp_build_context i32, h16, u8n; + LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type; + LLVMValueRef i32_c8, i32_c128, i32_c255; + LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi; + LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi; + LLVMValueRef x0, x1; + LLVMValueRef y0, y1; + LLVMValueRef neighbors[2][2]; + LLVMValueRef neighbors_lo[2][2]; + LLVMValueRef neighbors_hi[2][2]; + LLVMValueRef packed, packed_lo, packed_hi; + LLVMValueRef unswizzled[4]; + + lp_build_context_init(&i32, builder, lp_type_int(32)); + lp_build_context_init(&h16, builder, lp_type_ufixed(16)); + lp_build_context_init(&u8n, builder, lp_type_unorm(8)); + + i32_vec_type = lp_build_vec_type(i32.type); + h16_vec_type = lp_build_vec_type(h16.type); + u8n_vec_type = lp_build_vec_type(u8n.type); + + s = lp_build_mul_imm(&bld->coord_bld, s, 256); + t = lp_build_mul_imm(&bld->coord_bld, t, 256); + + s = LLVMBuildFPToSI(builder, s, i32_vec_type, ""); + t = LLVMBuildFPToSI(builder, t, i32_vec_type, ""); + + i32_c128 = lp_build_int_const_scalar(i32.type, -128); + s = LLVMBuildAdd(builder, s, i32_c128, ""); + t = LLVMBuildAdd(builder, t, i32_c128, ""); + + i32_c8 = lp_build_int_const_scalar(i32.type, 8); + s_ipart = LLVMBuildAShr(builder, s, i32_c8, ""); + t_ipart = LLVMBuildAShr(builder, t, i32_c8, ""); + + i32_c255 = lp_build_int_const_scalar(i32.type, 255); + s_fpart = LLVMBuildAnd(builder, s, i32_c255, ""); + t_fpart = LLVMBuildAnd(builder, t, i32_c255, ""); + + x0 = s_ipart; + y0 = t_ipart; + + x0 = lp_build_sample_wrap(bld, x0, width, bld->static_state->pot_width, bld->static_state->wrap_s); + y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t); + + x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one); + y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one); + + x1 = lp_build_sample_wrap(bld, x1, width, bld->static_state->pot_width, bld->static_state->wrap_s); + y1 = lp_build_sample_wrap(bld, y1, height, bld->static_state->pot_height, bld->static_state->wrap_t); + + /* + * Transform 4 x i32 in + * + * s_fpart = {s0, s1, s2, s3} + * + * into 8 x i16 + * + * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3} + * + * into two 8 x i16 + * + * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1} + * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3} + * + * and likewise for t_fpart. There is no risk of loosing precision here + * since the fractional parts only use the lower 8bits. + */ + + s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, ""); + t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, ""); + + { + LLVMTypeRef elem_type = LLVMInt32Type(); + LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH]; + LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH]; + LLVMValueRef shuffle_lo; + LLVMValueRef shuffle_hi; + unsigned i, j; + + for(j = 0; j < h16.type.length; j += 4) { + unsigned subindex = util_cpu_caps.little_endian ? 0 : 1; + LLVMValueRef index; + + index = LLVMConstInt(elem_type, j/2 + subindex, 0); + for(i = 0; i < 4; ++i) + shuffles_lo[j + i] = index; + + index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0); + for(i = 0; i < 4; ++i) + shuffles_hi[j + i] = index; + } + + shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length); + shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length); + + s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, ""); + t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, ""); + s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, ""); + t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, ""); + } + + /* + * Fetch the pixels as 4 x 32bit (rgba order might differ): + * + * rgba0 rgba1 rgba2 rgba3 + * + * bit cast them into 16 x u8 + * + * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3 + * + * unpack them into two 8 x i16: + * + * r0 g0 b0 a0 r1 g1 b1 a1 + * r2 g2 b2 a2 r3 g3 b3 a3 + * + * The higher 8 bits of the resulting elements will be zero. + */ + + neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_ptr); + neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_ptr); + neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_ptr); + neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_ptr); + + neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, ""); + neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, ""); + neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, ""); + neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, ""); + + lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]); + lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]); + lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]); + lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]); + + /* + * Linear interpolate with 8.8 fixed point. + */ + + packed_lo = lp_build_lerp_2d(&h16, + s_fpart_lo, t_fpart_lo, + neighbors_lo[0][0], + neighbors_lo[0][1], + neighbors_lo[1][0], + neighbors_lo[1][1]); + + packed_hi = lp_build_lerp_2d(&h16, + s_fpart_hi, t_fpart_hi, + neighbors_hi[0][0], + neighbors_hi[0][1], + neighbors_hi[1][0], + neighbors_hi[1][1]); + + packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi); + + /* + * Convert to SoA and swizzle. + */ + + packed = LLVMBuildBitCast(builder, packed, i32_vec_type, ""); + + lp_build_rgba8_to_f32_soa(bld->builder, + bld->texel_type, + packed, unswizzled); + + lp_build_format_swizzle_soa(bld->format_desc, + bld->texel_type, unswizzled, + texel); +} + + +static void +lp_build_sample_compare(struct lp_build_sample_context *bld, + LLVMValueRef p, + LLVMValueRef *texel) +{ + struct lp_build_context *texel_bld = &bld->texel_bld; + LLVMValueRef res; + unsigned chan; + + if(!bld->static_state->compare_mode) + return; + + /* TODO: Compare before swizzling, to avoid redundant computations */ + res = NULL; + for(chan = 0; chan < 4; ++chan) { + LLVMValueRef cmp; + cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]); + cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero); + + if(res) + res = lp_build_add(texel_bld, res, cmp); + else + res = cmp; + } + + assert(res); + res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25)); + + /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */ + for(chan = 0; chan < 3; ++chan) + texel[chan] = res; + texel[3] = texel_bld->one; +} + + +void +lp_build_sample_soa(LLVMBuilderRef builder, + const struct lp_sampler_static_state *static_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct lp_type type, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel) +{ + struct lp_build_sample_context bld; + LLVMValueRef width; + LLVMValueRef height; + LLVMValueRef stride; + LLVMValueRef data_ptr; + LLVMValueRef s; + LLVMValueRef t; + LLVMValueRef p; + + /* Setup our build context */ + memset(&bld, 0, sizeof bld); + bld.builder = builder; + bld.static_state = static_state; + bld.dynamic_state = dynamic_state; + bld.format_desc = util_format_description(static_state->format); + bld.coord_type = type; + bld.int_coord_type = lp_int_type(type); + bld.texel_type = type; + lp_build_context_init(&bld.coord_bld, builder, bld.coord_type); + lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type); + lp_build_context_init(&bld.texel_bld, builder, bld.texel_type); + + /* Get the dynamic state */ + width = dynamic_state->width(dynamic_state, builder, unit); + height = dynamic_state->height(dynamic_state, builder, unit); + stride = dynamic_state->stride(dynamic_state, builder, unit); + data_ptr = dynamic_state->data_ptr(dynamic_state, builder, unit); + + s = coords[0]; + t = coords[1]; + p = coords[2]; + + width = lp_build_broadcast_scalar(&bld.int_coord_bld, width); + height = lp_build_broadcast_scalar(&bld.int_coord_bld, height); + stride = lp_build_broadcast_scalar(&bld.int_coord_bld, stride); + + if(static_state->target == PIPE_TEXTURE_1D) + t = bld.coord_bld.zero; + + if(static_state->normalized_coords) { + LLVMTypeRef coord_vec_type = lp_build_vec_type(bld.coord_type); + LLVMValueRef fp_width = LLVMBuildSIToFP(builder, width, coord_vec_type, ""); + LLVMValueRef fp_height = LLVMBuildSIToFP(builder, height, coord_vec_type, ""); + s = lp_build_mul(&bld.coord_bld, s, fp_width); + t = lp_build_mul(&bld.coord_bld, t, fp_height); + } + + switch (static_state->min_img_filter) { + case PIPE_TEX_FILTER_NEAREST: + lp_build_sample_2d_nearest_soa(&bld, s, t, width, height, stride, data_ptr, texel); + break; + case PIPE_TEX_FILTER_LINEAR: + case PIPE_TEX_FILTER_ANISO: + if(lp_format_is_rgba8(bld.format_desc)) + lp_build_sample_2d_linear_aos(&bld, s, t, width, height, stride, data_ptr, texel); + else + lp_build_sample_2d_linear_soa(&bld, s, t, width, height, stride, data_ptr, texel); + break; + default: + assert(0); + } + + /* FIXME: respect static_state->min_mip_filter */; + /* FIXME: respect static_state->mag_img_filter */; + /* FIXME: respect static_state->prefilter */; + + lp_build_sample_compare(&bld, p, texel); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.c b/src/gallium/drivers/llvmpipe/lp_bld_struct.c index 14d2b10df9..3998ac374f 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_struct.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.c @@ -42,17 +42,30 @@ LLVMValueRef +lp_build_struct_get_ptr(LLVMBuilderRef builder, + LLVMValueRef ptr, + unsigned member, + const char *name) +{ + LLVMValueRef indices[2]; + LLVMValueRef member_ptr; + indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); + indices[1] = LLVMConstInt(LLVMInt32Type(), member, 0); + member_ptr = LLVMBuildGEP(builder, ptr, indices, Elements(indices), ""); + lp_build_name(member_ptr, "%s.%s_ptr", LLVMGetValueName(ptr), name); + return member_ptr; +} + + +LLVMValueRef lp_build_struct_get(LLVMBuilderRef builder, LLVMValueRef ptr, unsigned member, const char *name) { - LLVMValueRef indices[2]; LLVMValueRef member_ptr; LLVMValueRef res; - indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); - indices[1] = LLVMConstInt(LLVMInt32Type(), member, 0); - member_ptr = LLVMBuildGEP(builder, ptr, indices, Elements(indices), ""); + member_ptr = lp_build_struct_get_ptr(builder, ptr, member, name); res = LLVMBuildLoad(builder, member_ptr, ""); lp_build_name(res, "%s.%s", LLVMGetValueName(ptr), name); return res; diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.h b/src/gallium/drivers/llvmpipe/lp_bld_struct.h index cbefdc9f81..740392f561 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_struct.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.h @@ -53,6 +53,18 @@ offsetof(_ctype, _cmember)) +/** + * Get value pointer to a structure member. + */ +LLVMValueRef +lp_build_struct_get_ptr(LLVMBuilderRef builder, + LLVMValueRef ptr, + unsigned member, + const char *name); + +/** + * Get the value of a structure member. + */ LLVMValueRef lp_build_struct_get(LLVMBuilderRef builder, LLVMValueRef ptr, diff --git a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c index ac7eed9379..64e81f7b1f 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c @@ -64,7 +64,7 @@ LLVMValueRef lp_build_broadcast_scalar(struct lp_build_context *bld, LLVMValueRef scalar) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; LLVMValueRef res; unsigned i; @@ -83,7 +83,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld, LLVMValueRef a, unsigned channel) { - const union lp_type type = bld->type; + const struct lp_type type = bld->type; const unsigned n = type.length; unsigned i, j; @@ -115,7 +115,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld, * YY00 YY00 .... YY00 * YYYY YYYY .... YYYY <= output */ - union lp_type type4 = type; + struct lp_type type4 = type; const char shifts[4][2] = { { 1, 2}, {-1, 2}, @@ -161,7 +161,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld, LLVMValueRef lp_build_swizzle1_aos(struct lp_build_context *bld, LLVMValueRef a, - unsigned char swizzle[4]) + const unsigned char swizzle[4]) { const unsigned n = bld->type.length; unsigned i, j; @@ -192,7 +192,7 @@ LLVMValueRef lp_build_swizzle2_aos(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b, - unsigned char swizzle[4]) + const unsigned char swizzle[4]) { const unsigned n = bld->type.length; unsigned i, j; @@ -201,11 +201,12 @@ lp_build_swizzle2_aos(struct lp_build_context *bld, return lp_build_swizzle1_aos(bld, a, swizzle); if(a == b) { - swizzle[0] %= 4; - swizzle[1] %= 4; - swizzle[2] %= 4; - swizzle[3] %= 4; - return lp_build_swizzle1_aos(bld, a, swizzle); + unsigned char swizzle1[4]; + swizzle1[0] = swizzle[0] % 4; + swizzle1[1] = swizzle[1] % 4; + swizzle1[2] = swizzle[2] % 4; + swizzle1[3] = swizzle[3] % 4; + return lp_build_swizzle1_aos(bld, a, swizzle1); } if(swizzle[0] % 4 == 0 && diff --git a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h index d7dd6a8a60..b9472127a6 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h @@ -40,7 +40,7 @@ #include <llvm-c/Core.h> -union lp_type type; +struct lp_type; struct lp_build_context; @@ -73,7 +73,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld, LLVMValueRef lp_build_swizzle1_aos(struct lp_build_context *bld, LLVMValueRef a, - unsigned char swizzle[4]); + const unsigned char swizzle[4]); /** @@ -85,7 +85,7 @@ LLVMValueRef lp_build_swizzle2_aos(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b, - unsigned char swizzle[4]); + const unsigned char swizzle[4]); #endif /* !LP_BLD_SWIZZLE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h index 912db24aec..eddb7a83fa 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h @@ -39,31 +39,46 @@ struct tgsi_token; -union lp_type; +struct lp_type; struct lp_build_context; struct lp_build_mask_context; -typedef void -(*lp_emit_fetch_texel_soa_callback)( LLVMBuilderRef builder, - void *context, - unsigned unit, - unsigned num_coords, - const LLVMValueRef *coords, - LLVMValueRef lodbias, - LLVMValueRef *texel); +/** + * Sampler code generation interface. + * + * Although texture sampling is a requirement for TGSI translation, it is + * a very different problem with several different approaches to it. This + * structure establishes an interface for texture sampling code generation, so + * that we can easily use different texture sampling strategies. + */ +struct lp_build_sampler_soa +{ + void + (*destroy)( struct lp_build_sampler_soa *sampler ); + + void + (*emit_fetch_texel)( struct lp_build_sampler_soa *sampler, + LLVMBuilderRef builder, + struct lp_type type, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel); +}; + void lp_build_tgsi_soa(LLVMBuilderRef builder, const struct tgsi_token *tokens, - union lp_type type, + struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef consts_ptr, const LLVMValueRef *pos, const LLVMValueRef (*inputs)[4], LLVMValueRef (*outputs)[4], - lp_emit_fetch_texel_soa_callback emit_fetch_texel, - void *emit_fetch_texel_context); + struct lp_build_sampler_soa *sampler); #endif /* LP_BLD_TGSI_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c index f588bde983..94ce4ae831 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c @@ -78,6 +78,11 @@ #define CHAN_Z 2 #define CHAN_W 3 +#define QUAD_TOP_LEFT 0 +#define QUAD_TOP_RIGHT 1 +#define QUAD_BOTTOM_LEFT 2 +#define QUAD_BOTTOM_RIGHT 3 + struct lp_build_tgsi_soa_context { @@ -88,8 +93,7 @@ struct lp_build_tgsi_soa_context const LLVMValueRef (*inputs)[NUM_CHANNELS]; LLVMValueRef (*outputs)[NUM_CHANNELS]; - lp_emit_fetch_texel_soa_callback emit_fetch_texel; - void *emit_fetch_texel_context; + struct lp_build_sampler_soa *sampler; LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS]; LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS]; @@ -98,6 +102,51 @@ struct lp_build_tgsi_soa_context }; +static const unsigned char +swizzle_left[4] = { + QUAD_TOP_LEFT, QUAD_TOP_LEFT, + QUAD_BOTTOM_LEFT, QUAD_BOTTOM_LEFT +}; + +static const unsigned char +swizzle_right[4] = { + QUAD_TOP_RIGHT, QUAD_TOP_RIGHT, + QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT +}; + +static const unsigned char +swizzle_top[4] = { + QUAD_TOP_LEFT, QUAD_TOP_RIGHT, + QUAD_TOP_LEFT, QUAD_TOP_RIGHT +}; + +static const unsigned char +swizzle_bottom[4] = { + QUAD_BOTTOM_LEFT, QUAD_BOTTOM_RIGHT, + QUAD_BOTTOM_LEFT, QUAD_BOTTOM_RIGHT +}; + + +static LLVMValueRef +emit_ddx(struct lp_build_tgsi_soa_context *bld, + LLVMValueRef src) +{ + LLVMValueRef src_left = lp_build_swizzle1_aos(&bld->base, src, swizzle_left); + LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right); + return lp_build_sub(&bld->base, src_right, src_left); +} + + +static LLVMValueRef +emit_ddy(struct lp_build_tgsi_soa_context *bld, + LLVMValueRef src) +{ + LLVMValueRef src_top = lp_build_swizzle1_aos(&bld->base, src, swizzle_top); + LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom); + return lp_build_sub(&bld->base, src_top, src_bottom); +} + + /** * Register fetch. */ @@ -109,14 +158,14 @@ emit_fetch( const unsigned chan_index ) { const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[index]; - unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index ); + unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index ); LLVMValueRef res; switch (swizzle) { - case TGSI_EXTSWIZZLE_X: - case TGSI_EXTSWIZZLE_Y: - case TGSI_EXTSWIZZLE_Z: - case TGSI_EXTSWIZZLE_W: + case TGSI_SWIZZLE_X: + case TGSI_SWIZZLE_Y: + case TGSI_SWIZZLE_Z: + case TGSI_SWIZZLE_W: switch (reg->SrcRegister.File) { case TGSI_FILE_CONSTANT: { @@ -149,14 +198,6 @@ emit_fetch( } break; - case TGSI_EXTSWIZZLE_ZERO: - res = bld->base.zero; - break; - - case TGSI_EXTSWIZZLE_ONE: - res = bld->base.one; - break; - default: assert( 0 ); return bld->base.undef; @@ -168,6 +209,7 @@ emit_fetch( break; case TGSI_UTIL_SIGN_SET: + /* TODO: Use bitwese OR for floating point */ res = lp_build_abs( &bld->base, res ); res = LLVMBuildNeg( bld->base.builder, res, "" ); break; @@ -185,6 +227,36 @@ emit_fetch( /** + * Register fetch with derivatives. + */ +static void +emit_fetch_deriv( + struct lp_build_tgsi_soa_context *bld, + const struct tgsi_full_instruction *inst, + unsigned index, + const unsigned chan_index, + LLVMValueRef *res, + LLVMValueRef *ddx, + LLVMValueRef *ddy) +{ + LLVMValueRef src; + + src = emit_fetch(bld, inst, index, chan_index); + + if(res) + *res = src; + + /* TODO: use interpolation coeffs for inputs */ + + if(ddx) + *ddx = emit_ddx(bld, src); + + if(ddy) + *ddy = emit_ddy(bld, src); +} + + +/** * Register store. */ static void @@ -239,17 +311,18 @@ emit_store( * High-level instruction translators. */ + static void emit_tex( struct lp_build_tgsi_soa_context *bld, const struct tgsi_full_instruction *inst, boolean apply_lodbias, - boolean projected) + boolean projected, + LLVMValueRef *texel) { const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index; LLVMValueRef lodbias; LLVMValueRef oow; LLVMValueRef coords[3]; - LLVMValueRef texel[4]; unsigned num_coords; unsigned i; @@ -289,12 +362,11 @@ emit_tex( struct lp_build_tgsi_soa_context *bld, coords[i] = lp_build_mul(&bld->base, coords[i], oow); } - bld->emit_fetch_texel(bld->base.builder, bld->emit_fetch_texel_context, - unit, num_coords, coords, lodbias, texel); - - FOR_EACH_DST0_ENABLED_CHANNEL( inst, i ) { - emit_store( bld, inst, 0, i, texel[i] ); - } + bld->sampler->emit_fetch_texel(bld->sampler, + bld->base.builder, + bld->base.type, + unit, num_coords, coords, lodbias, + texel); } @@ -314,12 +386,7 @@ emit_kil( unsigned swizzle; /* Unswizzle channel */ - swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index ); - - /* Note that we test if the value is less than zero, so 1.0 and 0.0 need - * not to be tested. */ - if(swizzle == TGSI_EXTSWIZZLE_ZERO || swizzle == TGSI_EXTSWIZZLE_ONE) - continue; + swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index ); /* Check if the component has not been already tested. */ assert(swizzle < NUM_CHANNELS); @@ -347,14 +414,6 @@ emit_kil( } -static void -emit_kilp( - struct lp_build_tgsi_soa_context *bld ) -{ - /* XXX todo / fix me */ -} - - /** * Check if inst src/dest regs use indirect addressing into temporary * register file. @@ -382,45 +441,52 @@ indirect_temp_reference(const struct tgsi_full_instruction *inst) static int emit_instruction( struct lp_build_tgsi_soa_context *bld, - struct tgsi_full_instruction *inst ) + const struct tgsi_full_instruction *inst, + const struct tgsi_opcode_info *info) { unsigned chan_index; LLVMValueRef src0, src1, src2; LLVMValueRef tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; - LLVMValueRef dst0; + LLVMValueRef res; + LLVMValueRef dst0[NUM_CHANNELS]; /* we can't handle indirect addressing into temp register file yet */ if (indirect_temp_reference(inst)) return FALSE; + assert(info->num_dst <= 1); + if(info->num_dst) { + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + dst0[chan_index] = bld->base.undef; + } + } + switch (inst->Instruction.Opcode) { #if 0 case TGSI_OPCODE_ARL: + /* FIXME */ FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); emit_flr(bld, 0, 0); emit_f2it( bld, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; #endif case TGSI_OPCODE_MOV: - case TGSI_OPCODE_SWZ: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index ); } break; case TGSI_OPCODE_LIT: if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) { - emit_store( bld, inst, 0, CHAN_X, bld->base.one); + dst0[CHAN_X] = bld->base.one; } if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) { src0 = emit_fetch( bld, inst, 0, CHAN_X ); - dst0 = lp_build_max( &bld->base, src0, bld->base.zero); - emit_store( bld, inst, 0, CHAN_Y, dst0); + dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero); } if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) { /* XMM[1] = SrcReg[0].yyyy */ @@ -432,20 +498,19 @@ emit_instruction( tmp1 = lp_build_pow( &bld->base, tmp1, tmp2); tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero); - dst0 = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero); - emit_store( bld, inst, 0, CHAN_Z, dst0); + dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero); } if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) { - emit_store( bld, inst, 0, CHAN_W, bld->base.one); + dst0[CHAN_W] = bld->base.one; } break; case TGSI_OPCODE_RCP: /* TGSI_OPCODE_RECIP */ src0 = emit_fetch( bld, inst, 0, CHAN_X ); - dst0 = lp_build_rcp(&bld->base, src0); + res = lp_build_rcp(&bld->base, src0); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, dst0 ); + dst0[chan_index] = res; } break; @@ -453,9 +518,9 @@ emit_instruction( /* TGSI_OPCODE_RECIPSQRT */ src0 = emit_fetch( bld, inst, 0, CHAN_X ); src0 = lp_build_abs(&bld->base, src0); - dst0 = lp_build_rsqrt(&bld->base, src0); + res = lp_build_rsqrt(&bld->base, src0); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, dst0 ); + dst0[chan_index] = res; } break; @@ -479,16 +544,15 @@ emit_instruction( lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2); if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) - emit_store( bld, inst, 0, CHAN_X, tmp0); + dst0[CHAN_X] = tmp0; if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) - emit_store( bld, inst, 0, CHAN_Y, tmp1); + dst0[CHAN_Y] = tmp1; if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) - emit_store( bld, inst, 0, CHAN_Z, tmp2); + dst0[CHAN_Z] = tmp2; } /* dst.w = 1.0 */ if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) { - tmp0 = bld->base.one; - emit_store( bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = bld->base.one; } break; @@ -514,20 +578,18 @@ emit_instruction( /* dst.x = floor(lg2(abs(src.x))) */ if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) - emit_store( bld, inst, 0, CHAN_X, tmp0); + dst0[CHAN_X] = tmp0; /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */ if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) { - tmp1 = lp_build_div( &bld->base, src0, tmp1); - emit_store( bld, inst, 0, CHAN_Y, tmp1); + dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1); } /* dst.z = lg2(abs(src.x)) */ if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) - emit_store( bld, inst, 0, CHAN_Z, tmp2); + dst0[CHAN_Z] = tmp2; } /* dst.w = 1.0 */ if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) { - tmp0 = bld->base.one; - emit_store( bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = bld->base.one; } break; @@ -535,8 +597,7 @@ emit_instruction( FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); - dst0 = lp_build_mul(&bld->base, src0, src1); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_mul(&bld->base, src0, src1); } break; @@ -544,8 +605,7 @@ emit_instruction( FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); - dst0 = lp_build_add(&bld->base, src0, src1); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_add(&bld->base, src0, src1); } break; @@ -563,7 +623,7 @@ emit_instruction( tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); tmp0 = lp_build_add( &bld->base, tmp0, tmp1); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; @@ -585,28 +645,24 @@ emit_instruction( tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); tmp0 = lp_build_add( &bld->base, tmp0, tmp1); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; case TGSI_OPCODE_DST: IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) { - tmp0 = bld->base.one; - emit_store( bld, inst, 0, CHAN_X, tmp0); + dst0[CHAN_X] = bld->base.one; } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) { tmp0 = emit_fetch( bld, inst, 0, CHAN_Y ); tmp1 = emit_fetch( bld, inst, 1, CHAN_Y ); - tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); - emit_store( bld, inst, 0, CHAN_Y, tmp0); + dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1); } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { - tmp0 = emit_fetch( bld, inst, 0, CHAN_Z ); - emit_store( bld, inst, 0, CHAN_Z, tmp0); + dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z ); } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { - tmp0 = emit_fetch( bld, inst, 1, CHAN_W ); - emit_store( bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W ); } break; @@ -614,8 +670,7 @@ emit_instruction( FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); - dst0 = lp_build_min( &bld->base, src0, src1 ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_min( &bld->base, src0, src1 ); } break; @@ -623,8 +678,7 @@ emit_instruction( FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); - dst0 = lp_build_max( &bld->base, src0, src1 ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_max( &bld->base, src0, src1 ); } break; @@ -634,8 +688,7 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; @@ -645,8 +698,7 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; @@ -658,7 +710,7 @@ emit_instruction( tmp2 = emit_fetch( bld, inst, 2, chan_index ); tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); tmp0 = lp_build_add( &bld->base, tmp0, tmp2); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; @@ -666,8 +718,7 @@ emit_instruction( FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); tmp1 = emit_fetch( bld, inst, 1, chan_index ); - tmp0 = lp_build_sub( &bld->base, tmp0, tmp1); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1); } break; @@ -678,13 +729,19 @@ emit_instruction( src2 = emit_fetch( bld, inst, 2, chan_index ); tmp0 = lp_build_sub( &bld->base, src1, src2 ); tmp0 = lp_build_mul( &bld->base, src0, tmp0 ); - dst0 = lp_build_add( &bld->base, tmp0, src2 ); - emit_store( bld, inst, 0, chan_index, dst0 ); + dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 ); } break; case TGSI_OPCODE_CND: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + src2 = emit_fetch( bld, inst, 2, chan_index ); + tmp1 = lp_build_const_scalar(bld->base.type, 0.5); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 ); + } break; case TGSI_OPCODE_DP2A: @@ -698,45 +755,49 @@ emit_instruction( tmp1 = emit_fetch( bld, inst, 2, CHAN_X ); /* xmm1 = src[2].x */ tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */ FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); /* dest[ch] = xmm0 */ + dst0[chan_index] = tmp0; /* dest[ch] = xmm0 */ } break; -#if 0 case TGSI_OPCODE_FRC: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_frc( bld, 0, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + src0 = emit_fetch( bld, inst, 0, chan_index ); + tmp0 = lp_build_floor(&bld->base, src0); + tmp0 = lp_build_sub(&bld->base, tmp0, src0); + dst0[chan_index] = tmp0; } break; case TGSI_OPCODE_CLAMP: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + src2 = emit_fetch( bld, inst, 2, chan_index ); + tmp0 = lp_build_max(&bld->base, tmp0, src1); + tmp0 = lp_build_min(&bld->base, tmp0, src2); + dst0[chan_index] = tmp0; + } break; case TGSI_OPCODE_FLR: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_flr( bld, 0, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_floor(&bld->base, tmp0); } break; case TGSI_OPCODE_ROUND: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_rnd( bld, 0, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_round(&bld->base, tmp0); } break; -#endif case TGSI_OPCODE_EX2: { tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); tmp0 = lp_build_exp2( &bld->base, tmp0); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; } @@ -745,16 +806,16 @@ emit_instruction( tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); tmp0 = lp_build_log2( &bld->base, tmp0); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; case TGSI_OPCODE_POW: src0 = emit_fetch( bld, inst, 0, CHAN_X ); src1 = emit_fetch( bld, inst, 1, CHAN_X ); - dst0 = lp_build_pow( &bld->base, src0, src1 ); + res = lp_build_pow( &bld->base, src0, src1 ); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, dst0 ); + dst0[chan_index] = res; } break; @@ -775,7 +836,7 @@ emit_instruction( tmp5 = tmp3; tmp5 = lp_build_mul( &bld->base, tmp5, tmp4); tmp2 = lp_build_sub( &bld->base, tmp2, tmp5); - emit_store( bld, inst, 0, CHAN_X, tmp2); + dst0[CHAN_X] = tmp2; } if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) || IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) { @@ -786,31 +847,30 @@ emit_instruction( tmp3 = lp_build_mul( &bld->base, tmp3, tmp2); tmp1 = lp_build_mul( &bld->base, tmp1, tmp5); tmp3 = lp_build_sub( &bld->base, tmp3, tmp1); - emit_store( bld, inst, 0, CHAN_Y, tmp3); + dst0[CHAN_Y] = tmp3; } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { tmp5 = lp_build_mul( &bld->base, tmp5, tmp4); tmp0 = lp_build_mul( &bld->base, tmp0, tmp2); tmp5 = lp_build_sub( &bld->base, tmp5, tmp0); - emit_store( bld, inst, 0, CHAN_Z, tmp5); + dst0[CHAN_Z] = tmp5; } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { - tmp0 = bld->base.one; - emit_store( bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = bld->base.one; } break; case TGSI_OPCODE_ABS: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); - tmp0 = lp_build_abs( &bld->base, tmp0 ) ; - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_abs( &bld->base, tmp0 ); } break; case TGSI_OPCODE_RCC: + /* deprecated? */ + assert(0); return 0; - break; case TGSI_OPCODE_DPH: tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); @@ -827,7 +887,7 @@ emit_instruction( tmp1 = emit_fetch( bld, inst, 1, CHAN_W ); tmp0 = lp_build_add( &bld->base, tmp0, tmp1); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; @@ -835,25 +895,27 @@ emit_instruction( tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); tmp0 = lp_build_cos( &bld->base, tmp0 ); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; case TGSI_OPCODE_DDX: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL); + } break; case TGSI_OPCODE_DDY: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]); + } break; -#if 0 case TGSI_OPCODE_KILP: /* predicated kill */ - emit_kilp( bld ); - return 0; /* XXX fix me */ + /* FIXME */ + return 0; break; -#endif case TGSI_OPCODE_KIL: /* conditional kill */ @@ -885,13 +947,14 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; case TGSI_OPCODE_SFL: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + dst0[chan_index] = bld->base.zero; + } break; case TGSI_OPCODE_SGT: @@ -899,8 +962,7 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; @@ -908,7 +970,7 @@ emit_instruction( tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); tmp0 = lp_build_sin( &bld->base, tmp0 ); FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; @@ -917,8 +979,7 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; @@ -927,85 +988,99 @@ emit_instruction( src0 = emit_fetch( bld, inst, 0, chan_index ); src1 = emit_fetch( bld, inst, 1, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 ); - dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); } break; case TGSI_OPCODE_STR: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + dst0[chan_index] = bld->base.one; + } break; case TGSI_OPCODE_TEX: - emit_tex( bld, inst, FALSE, FALSE ); + emit_tex( bld, inst, FALSE, FALSE, dst0 ); break; case TGSI_OPCODE_TXD: + /* FIXME */ return 0; break; case TGSI_OPCODE_UP2H: + /* deprecated */ + assert (0); return 0; break; case TGSI_OPCODE_UP2US: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_UP4B: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_UP4UB: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_X2D: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_ARA: + /* deprecated */ + assert(0); return 0; break; #if 0 case TGSI_OPCODE_ARR: + /* FIXME */ FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); emit_rnd( bld, 0, 0 ); emit_f2it( bld, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = tmp0; } break; #endif case TGSI_OPCODE_BRA: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_CAL: + /* FIXME */ return 0; break; -#if 0 case TGSI_OPCODE_RET: - emit_ret( bld ); + /* FIXME */ + return 0; break; -#endif case TGSI_OPCODE_END: break; -#if 0 case TGSI_OPCODE_SSG: /* TGSI_OPCODE_SGN */ FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_sgn( bld, 0, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 ); } break; -#endif case TGSI_OPCODE_CMP: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { @@ -1013,34 +1088,29 @@ emit_instruction( src1 = emit_fetch( bld, inst, 1, chan_index ); src2 = emit_fetch( bld, inst, 2, chan_index ); tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero ); - dst0 = lp_build_select( &bld->base, tmp0, src1, src2); - emit_store( bld, inst, 0, chan_index, dst0); + dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2); } break; case TGSI_OPCODE_SCS: IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) { tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); - tmp0 = lp_build_cos( &bld->base, tmp0 ); - emit_store( bld, inst, 0, CHAN_X, tmp0); + dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 ); } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) { tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); - tmp0 = lp_build_sin( &bld->base, tmp0 ); - emit_store( bld, inst, 0, CHAN_Y, tmp0); + dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 ); } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { - tmp0 = bld->base.zero; - emit_store( bld, inst, 0, CHAN_Z, tmp0); + dst0[CHAN_Z] = bld->base.zero; } IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { - tmp0 = bld->base.one; - emit_store( bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = bld->base.one; } break; case TGSI_OPCODE_TXB: - emit_tex( bld, inst, TRUE, FALSE ); + emit_tex( bld, inst, TRUE, FALSE, dst0 ); break; case TGSI_OPCODE_NRM: @@ -1099,38 +1169,35 @@ emit_instruction( /* dst.x = xmm1 * src.x */ if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) { - tmp4 = lp_build_mul( &bld->base, tmp4, tmp1); - emit_store(bld, inst, 0, CHAN_X, tmp4); + dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1); } /* dst.y = xmm1 * src.y */ if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) { - tmp5 = lp_build_mul( &bld->base, tmp5, tmp1); - emit_store(bld, inst, 0, CHAN_Y, tmp5); + dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1); } /* dst.z = xmm1 * src.z */ if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) { - tmp6 = lp_build_mul( &bld->base, tmp6, tmp1); - emit_store(bld, inst, 0, CHAN_Z, tmp6); + dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1); } /* dst.w = xmm1 * src.w */ if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) { - tmp7 = lp_build_mul( &bld->base, tmp7, tmp1); - emit_store(bld, inst, 0, CHAN_W, tmp7); + dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1); } } - /* dst0.w = 1.0 */ + /* dst.w = 1.0 */ if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) { - tmp0 = bld->base.one; - emit_store(bld, inst, 0, CHAN_W, tmp0); + dst0[CHAN_W] = bld->base.one; } } break; case TGSI_OPCODE_DIV: + /* deprecated */ + assert( 0 ); return 0; break; @@ -1143,118 +1210,157 @@ emit_instruction( tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); /* xmm1 = xmm1 * xmm2 */ tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */ FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { - emit_store( bld, inst, 0, chan_index, tmp0); /* dest[ch] = xmm0 */ + dst0[chan_index] = tmp0; /* dest[ch] = xmm0 */ } break; case TGSI_OPCODE_TXL: - emit_tex( bld, inst, TRUE, FALSE ); + emit_tex( bld, inst, TRUE, FALSE, dst0 ); break; case TGSI_OPCODE_TXP: - emit_tex( bld, inst, FALSE, TRUE ); + emit_tex( bld, inst, FALSE, TRUE, dst0 ); break; case TGSI_OPCODE_BRK: + /* FIXME */ return 0; break; case TGSI_OPCODE_IF: + /* FIXME */ return 0; break; case TGSI_OPCODE_BGNFOR: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_REP: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_ELSE: + /* FIXME */ return 0; break; case TGSI_OPCODE_ENDIF: + /* FIXME */ return 0; break; case TGSI_OPCODE_ENDFOR: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_ENDREP: + /* deprecated */ + assert(0); return 0; break; case TGSI_OPCODE_PUSHA: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_POPA: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_CEIL: - return 0; + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + dst0[chan_index] = lp_build_ceil(&bld->base, tmp0); + } break; case TGSI_OPCODE_I2F: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_NOT: + /* deprecated? */ + assert(0); return 0; break; -#if 0 case TGSI_OPCODE_TRUNC: FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { tmp0 = emit_fetch( bld, inst, 0, chan_index ); - emit_f2it( bld, 0 ); - emit_i2f( bld, 0 ); - emit_store( bld, inst, 0, chan_index, tmp0); + dst0[chan_index] = lp_build_trunc(&bld->base, tmp0); } break; -#endif case TGSI_OPCODE_SHL: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_SHR: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_AND: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_OR: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_MOD: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_XOR: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_SAD: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_TXF: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_TXQ: + /* deprecated? */ + assert(0); return 0; break; case TGSI_OPCODE_CONT: + /* deprecated? */ + assert(0); return 0; break; @@ -1266,10 +1372,19 @@ emit_instruction( return 0; break; + case TGSI_OPCODE_NOP: + break; + default: return 0; } + if(info->num_dst) { + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, dst0[chan_index]); + } + } + return 1; } @@ -1277,14 +1392,13 @@ emit_instruction( void lp_build_tgsi_soa(LLVMBuilderRef builder, const struct tgsi_token *tokens, - union lp_type type, + struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef consts_ptr, const LLVMValueRef *pos, const LLVMValueRef (*inputs)[NUM_CHANNELS], LLVMValueRef (*outputs)[NUM_CHANNELS], - lp_emit_fetch_texel_soa_callback emit_fetch_texel, - void *emit_fetch_texel_context) + struct lp_build_sampler_soa *sampler) { struct lp_build_tgsi_soa_context bld; struct tgsi_parse_context parse; @@ -1299,8 +1413,7 @@ lp_build_tgsi_soa(LLVMBuilderRef builder, bld.inputs = inputs; bld.outputs = outputs; bld.consts_ptr = consts_ptr; - bld.emit_fetch_texel = emit_fetch_texel; - bld.emit_fetch_texel_context = emit_fetch_texel_context; + bld.sampler = sampler; tgsi_parse_init( &parse, tokens ); @@ -1309,16 +1422,18 @@ lp_build_tgsi_soa(LLVMBuilderRef builder, switch( parse.FullToken.Token.Type ) { case TGSI_TOKEN_TYPE_DECLARATION: - /* Input already interpolated */ + /* Inputs already interpolated */ break; case TGSI_TOKEN_TYPE_INSTRUCTION: - if (!emit_instruction( &bld, &parse.FullToken.FullInstruction )) { + { unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode; const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode); - _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n", - info ? info->mnemonic : "<invalid>"); - } + if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, info )) + _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n", + info ? info->mnemonic : "<invalid>"); + } + break; case TGSI_TOKEN_TYPE_IMMEDIATE: diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.c b/src/gallium/drivers/llvmpipe/lp_bld_type.c index 8e0026fd97..1320a26721 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_type.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_type.c @@ -33,7 +33,7 @@ LLVMTypeRef -lp_build_elem_type(union lp_type type) +lp_build_elem_type(struct lp_type type) { if (type.floating) { switch(type.width) { @@ -55,7 +55,7 @@ lp_build_elem_type(union lp_type type) LLVMTypeRef -lp_build_vec_type(union lp_type type) +lp_build_vec_type(struct lp_type type) { LLVMTypeRef elem_type = lp_build_elem_type(type); return LLVMVectorType(elem_type, type.length); @@ -69,7 +69,7 @@ lp_build_vec_type(union lp_type type) * type and check for identity. */ boolean -lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type) +lp_check_elem_type(struct lp_type type, LLVMTypeRef elem_type) { LLVMTypeKind elem_kind; @@ -107,7 +107,7 @@ lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type) boolean -lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type) +lp_check_vec_type(struct lp_type type, LLVMTypeRef vec_type) { LLVMTypeRef elem_type; @@ -128,7 +128,7 @@ lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type) boolean -lp_check_value(union lp_type type, LLVMValueRef val) +lp_check_value(struct lp_type type, LLVMValueRef val) { LLVMTypeRef vec_type; @@ -143,24 +143,55 @@ lp_check_value(union lp_type type, LLVMValueRef val) LLVMTypeRef -lp_build_int_elem_type(union lp_type type) +lp_build_int_elem_type(struct lp_type type) { return LLVMIntType(type.width); } LLVMTypeRef -lp_build_int_vec_type(union lp_type type) +lp_build_int_vec_type(struct lp_type type) { LLVMTypeRef elem_type = lp_build_int_elem_type(type); return LLVMVectorType(elem_type, type.length); } +struct lp_type +lp_int_type(struct lp_type type) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.width = type.width; + res_type.length = type.length; + + return res_type; +} + + +/** + * Return the type with twice the bit width (hence half the number of elements). + */ +struct lp_type +lp_wider_type(struct lp_type type) +{ + struct lp_type res_type; + + memcpy(&res_type, &type, sizeof res_type); + res_type.width *= 2; + res_type.length /= 2; + + assert(res_type.length); + + return res_type; +} + + void lp_build_context_init(struct lp_build_context *bld, LLVMBuilderRef builder, - union lp_type type) + struct lp_type type) { bld->builder = builder; bld->type = type; diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.h b/src/gallium/drivers/llvmpipe/lp_bld_type.h index 3ce566be64..2fb233d335 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_type.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_type.h @@ -43,71 +43,73 @@ /** + * Native SIMD register width. + * + * 128 for all architectures we care about. + */ +#define LP_NATIVE_VECTOR_WIDTH 128 + +/** * Several functions can only cope with vectors of length up to this value. * You may need to increase that value if you want to represent bigger vectors. */ #define LP_MAX_VECTOR_LENGTH 16 -#define LP_MAX_TYPE_WIDTH 64 - /** * The LLVM type system can't conveniently express all the things we care about * on the types used for intermediate computations, such as signed vs unsigned, * normalized values, or fixed point. */ -union lp_type { - struct { - /** - * Floating-point. Cannot be used with fixed. Integer numbers are - * represented by this zero. - */ - unsigned floating:1; - - /** - * Fixed-point. Cannot be used with floating. Integer numbers are - * represented by this zero. - */ - unsigned fixed:1; - - /** - * Whether it can represent negative values or not. - * - * If this is not set for floating point, it means that all values are - * assumed to be positive. - */ - unsigned sign:1; - - /** - * Whether values are normalized to fit [0, 1] interval, or [-1, 1] - * interval for signed types. - * - * For integer types it means the representable integer range should be - * interpreted as the interval above. - * - * For floating and fixed point formats it means the values should be - * clamped to the interval above. - */ - unsigned norm:1; - - /** - * Element width. - * - * For fixed point values, the fixed point is assumed to be at half the - * width. - */ - unsigned width:14; - - /** - * Vector length. - * - * width*length should be a power of two greater or equal to eight. - * - * @sa LP_MAX_VECTOR_LENGTH - */ - unsigned length:14; - }; - uint32_t value; +struct lp_type { + /** + * Floating-point. Cannot be used with fixed. Integer numbers are + * represented by this zero. + */ + unsigned floating:1; + + /** + * Fixed-point. Cannot be used with floating. Integer numbers are + * represented by this zero. + */ + unsigned fixed:1; + + /** + * Whether it can represent negative values or not. + * + * If this is not set for floating point, it means that all values are + * assumed to be positive. + */ + unsigned sign:1; + + /** + * Whether values are normalized to fit [0, 1] interval, or [-1, 1] + * interval for signed types. + * + * For integer types it means the representable integer range should be + * interpreted as the interval above. + * + * For floating and fixed point formats it means the values should be + * clamped to the interval above. + */ + unsigned norm:1; + + /** + * Element width. + * + * For fixed point values, the fixed point is assumed to be at half the + * width. + */ + unsigned width:14; + + /** + * Vector length. + * + * width*length should be a power of two greater or equal to eight. + * + * @sa LP_MAX_VECTOR_LENGTH + */ + unsigned length:14; }; @@ -124,7 +126,7 @@ struct lp_build_context * This not only describes the input/output LLVM types, but also whether * to normalize/clamp the results. */ - union lp_type type; + struct lp_type type; /** Same as lp_build_undef(type) */ LLVMValueRef undef; @@ -137,38 +139,131 @@ struct lp_build_context }; +static INLINE struct lp_type +lp_type_float(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.floating = TRUE; + res_type.sign = TRUE; + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + +static INLINE struct lp_type +lp_type_int(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.sign = TRUE; + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + +static INLINE struct lp_type +lp_type_uint(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + +static INLINE struct lp_type +lp_type_unorm(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.norm = TRUE; + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + +static INLINE struct lp_type +lp_type_fixed(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.sign = TRUE; + res_type.fixed = TRUE; + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + +static INLINE struct lp_type +lp_type_ufixed(unsigned width) +{ + struct lp_type res_type; + + memset(&res_type, 0, sizeof res_type); + res_type.fixed = TRUE; + res_type.width = width; + res_type.length = LP_NATIVE_VECTOR_WIDTH / width; + + return res_type; +} + + LLVMTypeRef -lp_build_elem_type(union lp_type type); +lp_build_elem_type(struct lp_type type); LLVMTypeRef -lp_build_vec_type(union lp_type type); +lp_build_vec_type(struct lp_type type); boolean -lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type); +lp_check_elem_type(struct lp_type type, LLVMTypeRef elem_type); boolean -lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type); +lp_check_vec_type(struct lp_type type, LLVMTypeRef vec_type); boolean -lp_check_value(union lp_type type, LLVMValueRef val); +lp_check_value(struct lp_type type, LLVMValueRef val); LLVMTypeRef -lp_build_int_elem_type(union lp_type type); +lp_build_int_elem_type(struct lp_type type); LLVMTypeRef -lp_build_int_vec_type(union lp_type type); +lp_build_int_vec_type(struct lp_type type); + + +struct lp_type +lp_int_type(struct lp_type type); + + +struct lp_type +lp_wider_type(struct lp_type type); void lp_build_context_init(struct lp_build_context *bld, LLVMBuilderRef builder, - union lp_type type); + struct lp_type type); #endif /* !LP_BLD_TYPE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_clear.c b/src/gallium/drivers/llvmpipe/lp_clear.c index 580cca5b46..bdcff94b9b 100644 --- a/src/gallium/drivers/llvmpipe/lp_clear.c +++ b/src/gallium/drivers/llvmpipe/lp_clear.c @@ -67,6 +67,7 @@ llvmpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba, util_pack_color(rgba, ps->format, &cv); lp_tile_cache_clear(llvmpipe->cbuf_cache[i], rgba, cv); } + llvmpipe->dirty_render_cache = TRUE; } if (buffers & PIPE_CLEAR_DEPTHSTENCIL) { diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c index 233d1df0e1..57e71f3e98 100644 --- a/src/gallium/drivers/llvmpipe/lp_context.c +++ b/src/gallium/drivers/llvmpipe/lp_context.c @@ -31,13 +31,13 @@ */ #include "draw/draw_context.h" +#include "draw/draw_vbuf.h" #include "pipe/p_defines.h" #include "util/u_math.h" #include "util/u_memory.h" #include "lp_clear.h" #include "lp_context.h" #include "lp_flush.h" -#include "lp_prim_setup.h" #include "lp_prim_vbuf.h" #include "lp_state.h" #include "lp_surface.h" @@ -107,11 +107,16 @@ static void llvmpipe_destroy( struct pipe_context *pipe ) if (llvmpipe->draw) draw_destroy( llvmpipe->draw ); - for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) + for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) { lp_destroy_tile_cache(llvmpipe->cbuf_cache[i]); + pipe_surface_reference(&llvmpipe->framebuffer.cbufs[i], NULL); + } + pipe_surface_reference(&llvmpipe->framebuffer.zsbuf, NULL); - for (i = 0; i < PIPE_MAX_SAMPLERS; i++) + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { lp_destroy_tex_tile_cache(llvmpipe->tex_cache[i]); + pipe_texture_reference(&llvmpipe->texture[i], NULL); + } for (i = 0; i < Elements(llvmpipe->constants); i++) { if (llvmpipe->constants[i].buffer) { @@ -141,8 +146,6 @@ llvmpipe_is_texture_referenced( struct pipe_context *pipe, return PIPE_REFERENCED_FOR_WRITE; } - /* FIXME: we also need to do the same for the texture cache */ - return PIPE_UNREFERENCED; } @@ -261,21 +264,21 @@ llvmpipe_create( struct pipe_screen *screen ) (struct tgsi_sampler **) llvmpipe->tgsi.vert_samplers_list); - llvmpipe->setup = lp_draw_render_stage(llvmpipe); - if (!llvmpipe->setup) - goto fail; - if (debug_get_bool_option( "LP_NO_RAST", FALSE )) llvmpipe->no_rast = TRUE; - if (debug_get_bool_option( "LP_NO_VBUF", FALSE )) { - /* Deprecated path -- vbuf is the intended interface to the draw module: - */ - draw_set_rasterize_stage(llvmpipe->draw, llvmpipe->setup); - } - else { - lp_init_vbuf(llvmpipe); - } + llvmpipe->vbuf_backend = lp_create_vbuf_backend(llvmpipe); + if (!llvmpipe->vbuf_backend) + goto fail; + + llvmpipe->vbuf = draw_vbuf_stage(llvmpipe->draw, llvmpipe->vbuf_backend); + if (!llvmpipe->vbuf) + goto fail; + + draw_set_rasterize_stage(llvmpipe->draw, llvmpipe->vbuf); + draw_set_render(llvmpipe->draw, llvmpipe->vbuf_backend); + + /* plug in AA line/point stages */ draw_install_aaline_stage(llvmpipe->draw, &llvmpipe->pipe); diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h index 8d5a0d4f1f..3ad95d0bfc 100644 --- a/src/gallium/drivers/llvmpipe/lp_context.h +++ b/src/gallium/drivers/llvmpipe/lp_context.h @@ -88,9 +88,6 @@ struct llvmpipe_context { /** Mapped vertex buffers */ ubyte *mapped_vbuffer[PIPE_MAX_ATTRIBS]; - /** Mapped constant buffers */ - void *mapped_constants[PIPE_SHADER_TYPES]; - /** Vertex format */ struct vertex_info vertex_info; struct vertex_info vertex_info_vbuf; @@ -124,9 +121,10 @@ struct llvmpipe_context { /** The primitive drawing context */ struct draw_context *draw; - struct draw_stage *setup; + + /** Draw module backend */ + struct vbuf_render *vbuf_backend; struct draw_stage *vbuf; - struct llvmpipe_vbuf_render *vbuf_render; boolean dirty_render_cache; diff --git a/src/gallium/drivers/llvmpipe/lp_draw_arrays.c b/src/gallium/drivers/llvmpipe/lp_draw_arrays.c index 89772e62d3..0aa13a1fc6 100644 --- a/src/gallium/drivers/llvmpipe/lp_draw_arrays.c +++ b/src/gallium/drivers/llvmpipe/lp_draw_arrays.c @@ -45,54 +45,6 @@ -static void -llvmpipe_map_constant_buffers(struct llvmpipe_context *lp) -{ - struct pipe_screen *screen = lp->pipe.screen; - uint i, size; - - for (i = 0; i < PIPE_SHADER_TYPES; i++) { - if (lp->constants[i].buffer && lp->constants[i].buffer->size) - lp->mapped_constants[i] = screen->buffer_map(screen, lp->constants[i].buffer, - PIPE_BUFFER_USAGE_CPU_READ); - } - - if (lp->constants[PIPE_SHADER_VERTEX].buffer) - size = lp->constants[PIPE_SHADER_VERTEX].buffer->size; - else - size = 0; - - lp->jit_context.constants = lp->mapped_constants[PIPE_SHADER_FRAGMENT]; - - draw_set_mapped_constant_buffer(lp->draw, - lp->mapped_constants[PIPE_SHADER_VERTEX], - size); -} - - -static void -llvmpipe_unmap_constant_buffers(struct llvmpipe_context *lp) -{ - struct pipe_screen *screen = lp->pipe.screen; - uint i; - - /* really need to flush all prims since the vert/frag shaders const buffers - * are going away now. - */ - draw_flush(lp->draw); - - draw_set_mapped_constant_buffer(lp->draw, NULL, 0); - - lp->jit_context.constants = NULL; - - for (i = 0; i < 2; i++) { - if (lp->constants[i].buffer && lp->constants[i].buffer->size) - screen->buffer_unmap(screen, lp->constants[i].buffer); - lp->mapped_constants[i] = NULL; - } -} - - boolean llvmpipe_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start, unsigned count) @@ -124,7 +76,6 @@ llvmpipe_draw_range_elements(struct pipe_context *pipe, llvmpipe_update_derived( lp ); llvmpipe_map_transfers(lp); - llvmpipe_map_constant_buffers(lp); /* * Map vertex buffers @@ -163,7 +114,6 @@ llvmpipe_draw_range_elements(struct pipe_context *pipe, /* Note: leave drawing surfaces mapped */ - llvmpipe_unmap_constant_buffers(lp); lp->dirty_render_cache = TRUE; diff --git a/src/gallium/drivers/llvmpipe/lp_flush.c b/src/gallium/drivers/llvmpipe/lp_flush.c index b5c1c95bb7..cd8381fe30 100644 --- a/src/gallium/drivers/llvmpipe/lp_flush.c +++ b/src/gallium/drivers/llvmpipe/lp_flush.c @@ -58,8 +58,10 @@ llvmpipe_flush( struct pipe_context *pipe, * in the hope that a later clear will wipe them out. */ for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) - if (llvmpipe->cbuf_cache[i]) + if (llvmpipe->cbuf_cache[i]) { + lp_tile_cache_map_transfers(llvmpipe->cbuf_cache[i]); lp_flush_tile_cache(llvmpipe->cbuf_cache[i]); + } /* Need this call for hardware buffers before swapbuffers. * @@ -71,8 +73,10 @@ llvmpipe_flush( struct pipe_context *pipe, } else if (flags & PIPE_FLUSH_RENDER_CACHE) { for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) - if (llvmpipe->cbuf_cache[i]) + if (llvmpipe->cbuf_cache[i]) { + lp_tile_cache_map_transfers(llvmpipe->cbuf_cache[i]); lp_flush_tile_cache(llvmpipe->cbuf_cache[i]); + } /* FIXME: untile zsbuf! */ diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c index d288460a1b..13535dd638 100644 --- a/src/gallium/drivers/llvmpipe/lp_jit.c +++ b/src/gallium/drivers/llvmpipe/lp_jit.c @@ -36,23 +36,57 @@ #include <llvm-c/Transforms/Scalar.h> #include "util/u_memory.h" +#include "util/u_cpu_detect.h" #include "lp_screen.h" #include "lp_bld_intr.h" +#include "lp_bld_misc.h" #include "lp_jit.h" static void lp_jit_init_globals(struct llvmpipe_screen *screen) { - /* struct lp_jit_context */ + LLVMTypeRef texture_type; + + /* struct lp_jit_texture */ { LLVMTypeRef elem_types[4]; + + elem_types[LP_JIT_TEXTURE_WIDTH] = LLVMInt32Type(); + elem_types[LP_JIT_TEXTURE_HEIGHT] = LLVMInt32Type(); + elem_types[LP_JIT_TEXTURE_STRIDE] = LLVMInt32Type(); + elem_types[LP_JIT_TEXTURE_DATA] = LLVMPointerType(LLVMInt8Type(), 0); + + texture_type = LLVMStructType(elem_types, Elements(elem_types), 0); + + LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, width, + screen->target, texture_type, + LP_JIT_TEXTURE_WIDTH); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, height, + screen->target, texture_type, + LP_JIT_TEXTURE_HEIGHT); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, stride, + screen->target, texture_type, + LP_JIT_TEXTURE_STRIDE); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, data, + screen->target, texture_type, + LP_JIT_TEXTURE_DATA); + LP_CHECK_STRUCT_SIZE(struct lp_jit_texture, + screen->target, texture_type); + + LLVMAddTypeName(screen->module, "texture", texture_type); + } + + /* struct lp_jit_context */ + { + LLVMTypeRef elem_types[5]; LLVMTypeRef context_type; elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* constants */ elem_types[1] = LLVMPointerType(LLVMInt8Type(), 0); /* samplers */ elem_types[2] = LLVMFloatType(); /* alpha_ref_value */ elem_types[3] = LLVMPointerType(LLVMInt8Type(), 0); /* blend_color */ + elem_types[4] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */ context_type = LLVMStructType(elem_types, Elements(elem_types), 0); @@ -64,6 +98,9 @@ lp_jit_init_globals(struct llvmpipe_screen *screen) screen->target, context_type, 2); LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, blend_color, screen->target, context_type, 3); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, textures, + screen->target, context_type, + LP_JIT_CONTEXT_TEXTURES_INDEX); LP_CHECK_STRUCT_SIZE(struct lp_jit_context, screen->target, context_type); @@ -112,12 +149,23 @@ lp_jit_screen_init(struct llvmpipe_screen *screen) { char *error = NULL; + util_cpu_detect(); + +#if 0 + /* For simulating less capable machines */ + util_cpu_caps.has_sse3 = 0; + util_cpu_caps.has_sse4_1 = 0; +#endif + + LLVMLinkInJIT(); + LLVMInitializeNativeTarget(); + screen->module = LLVMModuleCreateWithName("llvmpipe"); screen->provider = LLVMCreateModuleProviderForExistingModule(screen->module); if (LLVMCreateJITCompiler(&screen->engine, screen->provider, 1, &error)) { - fprintf(stderr, "%s\n", error); + _debug_printf("%s\n", error); LLVMDisposeMessage(error); abort(); } @@ -128,8 +176,15 @@ lp_jit_screen_init(struct llvmpipe_screen *screen) LLVMAddTargetData(screen->target, screen->pass); /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, * but there are more on SVN. */ + /* TODO: Add more passes */ LLVMAddConstantPropagationPass(screen->pass); - LLVMAddInstructionCombiningPass(screen->pass); + if(util_cpu_caps.has_sse4_1) { + /* FIXME: There is a bug in this pass, whereby the combination of fptosi + * and sitofp (necessary for trunc/floor/ceil/round implementation) + * somehow becomes invalid code. + */ + LLVMAddInstructionCombiningPass(screen->pass); + } LLVMAddPromoteMemoryToRegisterPass(screen->pass); LLVMAddGVNPass(screen->pass); LLVMAddCFGSimplificationPass(screen->pass); diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h index a7fb60f9f5..58f716ede2 100644 --- a/src/gallium/drivers/llvmpipe/lp_jit.h +++ b/src/gallium/drivers/llvmpipe/lp_jit.h @@ -38,11 +38,31 @@ #include "lp_bld_struct.h" +#include "pipe/p_state.h" + struct tgsi_sampler; struct llvmpipe_screen; +struct lp_jit_texture +{ + uint32_t width; + uint32_t height; + uint32_t stride; + const void *data; +}; + + +enum { + LP_JIT_TEXTURE_WIDTH = 0, + LP_JIT_TEXTURE_HEIGHT, + LP_JIT_TEXTURE_STRIDE, + LP_JIT_TEXTURE_DATA +}; + + + /** * This structure is passed directly to the generated fragment shader. * @@ -60,11 +80,12 @@ struct lp_jit_context struct tgsi_sampler **samplers; - /* TODO: alpha reference value */ float alpha_ref_value; - /* TODO: blend constant color */ + /* FIXME: store (also?) in floats */ uint8_t *blend_color; + + struct lp_jit_texture textures[PIPE_MAX_SAMPLERS]; }; @@ -80,6 +101,11 @@ struct lp_jit_context #define lp_jit_context_blend_color(_builder, _ptr) \ lp_build_struct_get(_builder, _ptr, 3, "blend_color") +#define LP_JIT_CONTEXT_TEXTURES_INDEX 4 + +#define lp_jit_context_textures(_builder, _ptr) \ + lp_build_struct_get_ptr(_builder, _ptr, LP_JIT_CONTEXT_TEXTURES_INDEX, "textures") + typedef void (*lp_jit_frag_func)(struct lp_jit_context *context, diff --git a/src/gallium/drivers/llvmpipe/lp_prim_setup.c b/src/gallium/drivers/llvmpipe/lp_prim_setup.c deleted file mode 100644 index b14f8fb99d..0000000000 --- a/src/gallium/drivers/llvmpipe/lp_prim_setup.c +++ /dev/null @@ -1,190 +0,0 @@ -/************************************************************************** - * - * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. - * All Rights Reserved. - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the - * "Software"), to deal in the Software without restriction, including - * without limitation the rights to use, copy, modify, merge, publish, - * distribute, sub license, and/or sell copies of the Software, and to - * permit persons to whom the Software is furnished to do so, subject to - * the following conditions: - * - * The above copyright notice and this permission notice (including the - * next paragraph) shall be included in all copies or substantial portions - * of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS - * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF - * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. - * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR - * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, - * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE - * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. - * - **************************************************************************/ - -/** - * \brief A draw stage that drives our triangle setup routines from - * within the draw pipeline. One of two ways to drive setup, the - * other being in lp_prim_vbuf.c. - * - * \author Keith Whitwell <keith@tungstengraphics.com> - * \author Brian Paul - */ - - -#include "lp_context.h" -#include "lp_setup.h" -#include "lp_state.h" -#include "lp_prim_setup.h" -#include "draw/draw_pipe.h" -#include "draw/draw_vertex.h" -#include "util/u_memory.h" - -/** - * Triangle setup info (derived from draw_stage). - * Also used for line drawing (taking some liberties). - */ -struct setup_stage { - struct draw_stage stage; /**< This must be first (base class) */ - - struct setup_context *setup; -}; - - - -/** - * Basically a cast wrapper. - */ -static INLINE struct setup_stage *setup_stage( struct draw_stage *stage ) -{ - return (struct setup_stage *)stage; -} - - -typedef const float (*cptrf4)[4]; - -static void -do_tri(struct draw_stage *stage, struct prim_header *prim) -{ - struct setup_stage *setup = setup_stage( stage ); - - llvmpipe_setup_tri( setup->setup, - (cptrf4)prim->v[0]->data, - (cptrf4)prim->v[1]->data, - (cptrf4)prim->v[2]->data ); -} - -static void -do_line(struct draw_stage *stage, struct prim_header *prim) -{ - struct setup_stage *setup = setup_stage( stage ); - - llvmpipe_setup_line( setup->setup, - (cptrf4)prim->v[0]->data, - (cptrf4)prim->v[1]->data ); -} - -static void -do_point(struct draw_stage *stage, struct prim_header *prim) -{ - struct setup_stage *setup = setup_stage( stage ); - - llvmpipe_setup_point( setup->setup, - (cptrf4)prim->v[0]->data ); -} - - - - -static void setup_begin( struct draw_stage *stage ) -{ - struct setup_stage *setup = setup_stage(stage); - - llvmpipe_setup_prepare( setup->setup ); - - stage->point = do_point; - stage->line = do_line; - stage->tri = do_tri; -} - - -static void setup_first_point( struct draw_stage *stage, - struct prim_header *header ) -{ - setup_begin(stage); - stage->point( stage, header ); -} - -static void setup_first_line( struct draw_stage *stage, - struct prim_header *header ) -{ - setup_begin(stage); - stage->line( stage, header ); -} - - -static void setup_first_tri( struct draw_stage *stage, - struct prim_header *header ) -{ - setup_begin(stage); - stage->tri( stage, header ); -} - - - -static void setup_flush( struct draw_stage *stage, - unsigned flags ) -{ - stage->point = setup_first_point; - stage->line = setup_first_line; - stage->tri = setup_first_tri; -} - - -static void reset_stipple_counter( struct draw_stage *stage ) -{ -} - - -static void render_destroy( struct draw_stage *stage ) -{ - struct setup_stage *ssetup = setup_stage(stage); - llvmpipe_setup_destroy_context(ssetup->setup); - FREE( stage ); -} - - -/** - * Create a new primitive setup/render stage. - */ -struct draw_stage *lp_draw_render_stage( struct llvmpipe_context *llvmpipe ) -{ - struct setup_stage *sstage = CALLOC_STRUCT(setup_stage); - - sstage->setup = llvmpipe_setup_create_context(llvmpipe); - sstage->stage.draw = llvmpipe->draw; - sstage->stage.point = setup_first_point; - sstage->stage.line = setup_first_line; - sstage->stage.tri = setup_first_tri; - sstage->stage.flush = setup_flush; - sstage->stage.reset_stipple_counter = reset_stipple_counter; - sstage->stage.destroy = render_destroy; - - return (struct draw_stage *)sstage; -} - -struct setup_context * -lp_draw_setup_context( struct draw_stage *stage ) -{ - struct setup_stage *ssetup = setup_stage(stage); - return ssetup->setup; -} - -void -lp_draw_flush( struct draw_stage *stage ) -{ - stage->flush( stage, 0 ); -} diff --git a/src/gallium/drivers/llvmpipe/lp_prim_setup.h b/src/gallium/drivers/llvmpipe/lp_prim_setup.h deleted file mode 100644 index da6cae6375..0000000000 --- a/src/gallium/drivers/llvmpipe/lp_prim_setup.h +++ /dev/null @@ -1,85 +0,0 @@ -/************************************************************************** - * - * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. - * All Rights Reserved. - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the - * "Software"), to deal in the Software without restriction, including - * without limitation the rights to use, copy, modify, merge, publish, - * distribute, sub license, and/or sell copies of the Software, and to - * permit persons to whom the Software is furnished to do so, subject to - * the following conditions: - * - * The above copyright notice and this permission notice (including the - * next paragraph) shall be included in all copies or substantial portions - * of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS - * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF - * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. - * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR - * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, - * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE - * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. - * - **************************************************************************/ - - -#ifndef LP_PRIM_SETUP_H -#define LP_PRIM_SETUP_H - - -/** - * vbuf is a special stage to gather the stream of triangles, lines, points - * together and reconstruct vertex buffers for hardware upload. - * - * First attempt, work in progress. - * - * TODO: - * - separate out vertex buffer building and primitive emit, ie >1 draw per vb. - * - tell vbuf stage how to build hw vertices directly - * - pass vbuf stage a buffer pointer for direct emit to agp/vram. - * - * - * - * Vertices are just an array of floats, with all the attributes - * packed. We currently assume a layout like: - * - * attr[0][0..3] - window position - * attr[1..n][0..3] - remaining attributes. - * - * Attributes are assumed to be 4 floats wide but are packed so that - * all the enabled attributes run contiguously. - */ - - -struct draw_stage; -struct llvmpipe_context; - - -typedef void (*vbuf_draw_func)( struct pipe_context *pipe, - unsigned prim, - const ushort *elements, - unsigned nr_elements, - const void *vertex_buffer, - unsigned nr_vertices ); - - -extern struct draw_stage * -lp_draw_render_stage( struct llvmpipe_context *llvmpipe ); - -extern struct setup_context * -lp_draw_setup_context( struct draw_stage * ); - -extern void -lp_draw_flush( struct draw_stage * ); - - -extern struct draw_stage * -lp_draw_vbuf_stage( struct draw_context *draw_context, - struct pipe_context *pipe, - vbuf_draw_func draw ); - - -#endif /* LP_PRIM_SETUP_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c index c394dcb61d..4abff4eccc 100644 --- a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c +++ b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c @@ -37,10 +37,9 @@ #include "lp_context.h" +#include "lp_setup.h" #include "lp_state.h" #include "lp_prim_vbuf.h" -#include "lp_prim_setup.h" -#include "lp_setup.h" #include "draw/draw_context.h" #include "draw/draw_vbuf.h" #include "util/u_memory.h" @@ -59,6 +58,8 @@ struct llvmpipe_vbuf_render { struct vbuf_render base; struct llvmpipe_context *llvmpipe; + struct setup_context *setup; + uint prim; uint vertex_size; uint nr_vertices; @@ -75,6 +76,11 @@ llvmpipe_vbuf_render(struct vbuf_render *vbr) } + + + + + static const struct vertex_info * lp_vbuf_get_vertex_info(struct vbuf_render *vbr) { @@ -105,36 +111,6 @@ lp_vbuf_allocate_vertices(struct vbuf_render *vbr, static void lp_vbuf_release_vertices(struct vbuf_render *vbr) { -#if 0 - { - struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); - const struct vertex_info *info = - llvmpipe_get_vbuf_vertex_info(cvbr->llvmpipe); - const float *vtx = (const float *) cvbr->vertex_buffer; - uint i, j; - debug_printf("%s (vtx_size = %u, vtx_used = %u)\n", - __FUNCTION__, cvbr->vertex_size, cvbr->nr_vertices); - for (i = 0; i < cvbr->nr_vertices; i++) { - for (j = 0; j < info->num_attribs; j++) { - uint k; - switch (info->attrib[j].emit) { - case EMIT_4F: k = 4; break; - case EMIT_3F: k = 3; break; - case EMIT_2F: k = 2; break; - case EMIT_1F: k = 1; break; - default: assert(0); - } - debug_printf("Vert %u attr %u: ", i, j); - while (k-- > 0) { - debug_printf("%g ", vtx[0]); - vtx++; - } - debug_printf("\n"); - } - } - } -#endif - /* keep the old allocation for next time */ } @@ -160,11 +136,7 @@ static boolean lp_vbuf_set_primitive(struct vbuf_render *vbr, unsigned prim) { struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); - - /* XXX: break this dependency - make setup_context live under - * llvmpipe, rename the old "setup" draw stage to something else. - */ - struct setup_context *setup_ctx = lp_draw_setup_context(cvbr->llvmpipe->setup); + struct setup_context *setup_ctx = cvbr->setup; llvmpipe_setup_prepare( setup_ctx ); @@ -193,14 +165,9 @@ lp_vbuf_draw(struct vbuf_render *vbr, const ushort *indices, uint nr) struct llvmpipe_context *llvmpipe = cvbr->llvmpipe; const unsigned stride = llvmpipe->vertex_info_vbuf.size * sizeof(float); const void *vertex_buffer = cvbr->vertex_buffer; + struct setup_context *setup_ctx = cvbr->setup; unsigned i; - /* XXX: break this dependency - make setup_context live under - * llvmpipe, rename the old "setup" draw stage to something else. - */ - struct draw_stage *setup = llvmpipe->setup; - struct setup_context *setup_ctx = lp_draw_setup_context(setup); - switch (cvbr->prim) { case PIPE_PRIM_POINTS: for (i = 0; i < nr; i++) { @@ -367,11 +334,6 @@ lp_vbuf_draw(struct vbuf_render *vbr, const ushort *indices, uint nr) default: assert(0); } - - /* XXX: why are we calling this??? If we had to call something, it - * would be a function in lp_setup.c: - */ - lp_draw_flush( setup ); } @@ -384,17 +346,12 @@ lp_vbuf_draw_arrays(struct vbuf_render *vbr, uint start, uint nr) { struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); struct llvmpipe_context *llvmpipe = cvbr->llvmpipe; + struct setup_context *setup_ctx = cvbr->setup; const unsigned stride = llvmpipe->vertex_info_vbuf.size * sizeof(float); const void *vertex_buffer = (void *) get_vert(cvbr->vertex_buffer, start, stride); unsigned i; - /* XXX: break this dependency - make setup_context live under - * llvmpipe, rename the old "setup" draw stage to something else. - */ - struct draw_stage *setup = llvmpipe->setup; - struct setup_context *setup_ctx = lp_draw_setup_context(setup); - switch (cvbr->prim) { case PIPE_PRIM_POINTS: for (i = 0; i < nr; i++) { @@ -568,40 +525,38 @@ static void lp_vbuf_destroy(struct vbuf_render *vbr) { struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); - cvbr->llvmpipe->vbuf_render = NULL; + llvmpipe_setup_destroy_context(cvbr->setup); FREE(cvbr); } /** - * Initialize the post-transform vertex buffer information for the given - * context. + * Create the post-transform vertex handler for the given context. */ -void -lp_init_vbuf(struct llvmpipe_context *lp) +struct vbuf_render * +lp_create_vbuf_backend(struct llvmpipe_context *lp) { - assert(lp->draw); + struct llvmpipe_vbuf_render *cvbr = CALLOC_STRUCT(llvmpipe_vbuf_render); - lp->vbuf_render = CALLOC_STRUCT(llvmpipe_vbuf_render); + assert(lp->draw); - lp->vbuf_render->base.max_indices = LP_MAX_VBUF_INDEXES; - lp->vbuf_render->base.max_vertex_buffer_bytes = LP_MAX_VBUF_SIZE; - lp->vbuf_render->base.get_vertex_info = lp_vbuf_get_vertex_info; - lp->vbuf_render->base.allocate_vertices = lp_vbuf_allocate_vertices; - lp->vbuf_render->base.map_vertices = lp_vbuf_map_vertices; - lp->vbuf_render->base.unmap_vertices = lp_vbuf_unmap_vertices; - lp->vbuf_render->base.set_primitive = lp_vbuf_set_primitive; - lp->vbuf_render->base.draw = lp_vbuf_draw; - lp->vbuf_render->base.draw_arrays = lp_vbuf_draw_arrays; - lp->vbuf_render->base.release_vertices = lp_vbuf_release_vertices; - lp->vbuf_render->base.destroy = lp_vbuf_destroy; + cvbr->base.max_indices = LP_MAX_VBUF_INDEXES; + cvbr->base.max_vertex_buffer_bytes = LP_MAX_VBUF_SIZE; - lp->vbuf_render->llvmpipe = lp; + cvbr->base.get_vertex_info = lp_vbuf_get_vertex_info; + cvbr->base.allocate_vertices = lp_vbuf_allocate_vertices; + cvbr->base.map_vertices = lp_vbuf_map_vertices; + cvbr->base.unmap_vertices = lp_vbuf_unmap_vertices; + cvbr->base.set_primitive = lp_vbuf_set_primitive; + cvbr->base.draw = lp_vbuf_draw; + cvbr->base.draw_arrays = lp_vbuf_draw_arrays; + cvbr->base.release_vertices = lp_vbuf_release_vertices; + cvbr->base.destroy = lp_vbuf_destroy; - lp->vbuf = draw_vbuf_stage(lp->draw, &lp->vbuf_render->base); + cvbr->llvmpipe = lp; - draw_set_rasterize_stage(lp->draw, lp->vbuf); + cvbr->setup = llvmpipe_setup_create_context(cvbr->llvmpipe); - draw_set_render(lp->draw, &lp->vbuf_render->base); + return &cvbr->base; } diff --git a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h index 6c4e6063e6..0676e2f42a 100644 --- a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h +++ b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h @@ -31,8 +31,8 @@ struct llvmpipe_context; -extern void -lp_init_vbuf(struct llvmpipe_context *llvmpipe); +extern struct vbuf_render * +lp_create_vbuf_backend(struct llvmpipe_context *llvmpipe); #endif /* LP_VBUF_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c index 125035771e..0518927458 100644 --- a/src/gallium/drivers/llvmpipe/lp_screen.c +++ b/src/gallium/drivers/llvmpipe/lp_screen.c @@ -27,8 +27,6 @@ #include "util/u_memory.h" -#include "util/u_simple_screen.h" -#include "pipe/internal/p_winsys_screen.h" #include "pipe/p_defines.h" #include "pipe/p_screen.h" @@ -67,8 +65,6 @@ llvmpipe_get_param(struct pipe_screen *screen, int param) return 1; case PIPE_CAP_GLSL: return 1; - case PIPE_CAP_S3TC: - return 0; case PIPE_CAP_ANISOTROPIC_FILTER: return 0; case PIPE_CAP_POINT_SPRITE: @@ -86,7 +82,7 @@ llvmpipe_get_param(struct pipe_screen *screen, int param) case PIPE_CAP_MAX_TEXTURE_2D_LEVELS: return 13; /* max 4Kx4K */ case PIPE_CAP_MAX_TEXTURE_3D_LEVELS: - return 8; /* max 128x128x128 */ + return 9; /* max 256x256x256 */ case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS: return 13; /* max 4Kx4K */ case PIPE_CAP_TGSI_CONT_SUPPORTED: @@ -196,8 +192,7 @@ static void llvmpipe_destroy_screen( struct pipe_screen *_screen ) { struct llvmpipe_screen *screen = llvmpipe_screen(_screen); - - struct pipe_winsys *winsys = _screen->winsys; + struct llvmpipe_winsys *winsys = screen->winsys; lp_jit_screen_cleanup(screen); diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c index d145f6d6bb..ffcbc9a379 100644 --- a/src/gallium/drivers/llvmpipe/lp_setup.c +++ b/src/gallium/drivers/llvmpipe/lp_setup.c @@ -33,7 +33,6 @@ */ #include "lp_context.h" -#include "lp_prim_setup.h" #include "lp_quad.h" #include "lp_setup.h" #include "lp_state.h" @@ -44,6 +43,7 @@ #include "pipe/p_thread.h" #include "util/u_math.h" #include "util/u_memory.h" +#include "lp_bld_debug.h" #include "lp_tile_cache.h" #include "lp_tile_soa.h" @@ -89,6 +89,8 @@ struct setup_context { float oneoverarea; int facing; + float pixel_offset; + struct quad_header quad[MAX_QUADS]; struct quad_header *quad_ptrs[MAX_QUADS]; unsigned count; @@ -114,6 +116,7 @@ struct setup_context { /** * Execute fragment shader for the four fragments in the quad. */ +ALIGN_STACK static void shade_quads(struct llvmpipe_context *llvmpipe, struct quad_header *quads[], @@ -123,7 +126,7 @@ shade_quads(struct llvmpipe_context *llvmpipe, struct quad_header *quad = quads[0]; const unsigned x = quad->input.x0; const unsigned y = quad->input.y0; - uint8_t *tile = lp_get_cached_tile(llvmpipe->cbuf_cache[0], x, y); + uint8_t *tile; uint8_t *color; void *depth; uint32_t ALIGN16_ATTRIB mask[4][NUM_CHANNELS]; @@ -149,7 +152,13 @@ shade_quads(struct llvmpipe_context *llvmpipe, mask[q][chan_index] = quads[q]->inout.mask & (1 << chan_index) ? ~0 : 0; /* color buffer */ - color = &TILE_PIXEL(tile, x & (TILE_SIZE-1), y & (TILE_SIZE-1), 0); + if(llvmpipe->framebuffer.nr_cbufs >= 1 && + llvmpipe->framebuffer.cbufs[0]) { + tile = lp_get_cached_tile(llvmpipe->cbuf_cache[0], x, y); + color = &TILE_PIXEL(tile, x & (TILE_SIZE-1), y & (TILE_SIZE-1), 0); + } + else + color = NULL; /* depth buffer */ if(llvmpipe->zsbuf_map) { @@ -162,12 +171,12 @@ shade_quads(struct llvmpipe_context *llvmpipe, else depth = NULL; - /* TODO: blend color */ + /* XXX: This will most likely fail on 32bit x86 without -mstackrealign */ + assert(lp_check_alignment(mask, 16)); - assert((((uintptr_t)mask) & 0xf) == 0); - assert((((uintptr_t)depth) & 0xf) == 0); - assert((((uintptr_t)color) & 0xf) == 0); - assert((((uintptr_t)llvmpipe->jit_context.blend_color) & 0xf) == 0); + assert(lp_check_alignment(depth, 16)); + assert(lp_check_alignment(color, 16)); + assert(lp_check_alignment(llvmpipe->jit_context.blend_color, 16)); /* run shader */ fs->current->jit_function( &llvmpipe->jit_context, @@ -270,11 +279,13 @@ clip_emit_quad( struct setup_context *setup, struct quad_header *quad ) * until we codegenerate single-quad variants of the fragment pipeline * we need this hack. */ const unsigned nr_quads = TILE_VECTOR_HEIGHT*TILE_VECTOR_WIDTH/QUAD_SIZE; - struct quad_header quads[nr_quads]; - struct quad_header *quad_ptrs[nr_quads]; + struct quad_header quads[4]; + struct quad_header *quad_ptrs[4]; int x0 = block_x(quad->input.x0); unsigned i; + assert(nr_quads == 4); + for(i = 0; i < nr_quads; ++i) { int x = x0 + 2*i; if(x == quad->input.x0) @@ -473,6 +484,16 @@ static boolean setup_sort_vertices( struct setup_context *setup, ((det > 0.0) ^ (setup->llvmpipe->rasterizer->front_winding == PIPE_WINDING_CW)); + /* Prepare pixel offset for rasterisation: + * - pixel center (0.5, 0.5) for GL, or + * - assume (0.0, 0.0) for other APIs. + */ + if (setup->llvmpipe->rasterizer->gl_rasterization_rules) { + setup->pixel_offset = 0.5f; + } else { + setup->pixel_offset = 0.0f; + } + return TRUE; } @@ -498,7 +519,7 @@ static void tri_pos_coeff( struct setup_context *setup, /* calculate a0 as the value which would be sampled for the * fragment at (0,0), taking into account that we want to sample at - * pixel centers, in other words (0.5, 0.5). + * pixel centers, in other words (pixel_offset, pixel_offset). * * this is neat but unfortunately not a good way to do things for * triangles with very large values of dadx or dady as it will @@ -509,8 +530,8 @@ static void tri_pos_coeff( struct setup_context *setup, * instead - i'll switch to this later. */ setup->coef.a0[0][i] = (setup->vmin[vertSlot][i] - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); /* debug_printf("attr[%d].%c: %f dx:%f dy:%f\n", @@ -599,8 +620,8 @@ static void tri_linear_coeff( struct setup_context *setup, * instead - i'll switch to this later. */ setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); /* debug_printf("attr[%d].%c: %f dx:%f dy:%f\n", @@ -651,8 +672,8 @@ static void tri_persp_coeff( struct setup_context *setup, setup->coef.dadx[1 + attrib][i] = dadx; setup->coef.dady[1 + attrib][i] = dady; setup->coef.a0[1 + attrib][i] = (mina - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); } } @@ -736,12 +757,12 @@ static void setup_tri_coefficients( struct setup_context *setup ) static void setup_tri_edges( struct setup_context *setup ) { - float vmin_x = setup->vmin[0][0] + 0.5f; - float vmid_x = setup->vmid[0][0] + 0.5f; + float vmin_x = setup->vmin[0][0] + setup->pixel_offset; + float vmid_x = setup->vmid[0][0] + setup->pixel_offset; - float vmin_y = setup->vmin[0][1] - 0.5f; - float vmid_y = setup->vmid[0][1] - 0.5f; - float vmax_y = setup->vmax[0][1] - 0.5f; + float vmin_y = setup->vmin[0][1] - setup->pixel_offset; + float vmid_y = setup->vmid[0][1] - setup->pixel_offset; + float vmax_y = setup->vmax[0][1] - setup->pixel_offset; setup->emaj.sy = ceilf(vmin_y); setup->emaj.lines = (int) ceilf(vmax_y - setup->emaj.sy); @@ -940,8 +961,8 @@ linear_pos_coeff(struct setup_context *setup, setup->coef.dadx[0][i] = dadx; setup->coef.dady[0][i] = dady; setup->coef.a0[0][i] = (setup->vmin[vertSlot][i] - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); } @@ -962,8 +983,8 @@ line_linear_coeff(struct setup_context *setup, setup->coef.dadx[1 + attrib][i] = dadx; setup->coef.dady[1 + attrib][i] = dady; setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); } } @@ -988,8 +1009,8 @@ line_persp_coeff(struct setup_context *setup, setup->coef.dadx[1 + attrib][i] = dadx; setup->coef.dady[1 + attrib][i] = dady; setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - - (dadx * (setup->vmin[0][0] - 0.5f) + - dady * (setup->vmin[0][1] - 0.5f))); + (dadx * (setup->vmin[0][0] - setup->pixel_offset) + + dady * (setup->vmin[0][1] - setup->pixel_offset))); } } diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h index fb10329887..7b26ce61a3 100644 --- a/src/gallium/drivers/llvmpipe/lp_state.h +++ b/src/gallium/drivers/llvmpipe/lp_state.h @@ -36,6 +36,7 @@ #include "pipe/p_state.h" #include "tgsi/tgsi_scan.h" #include "lp_jit.h" +#include "lp_bld_sample.h" /* for struct lp_sampler_static_state */ #define LP_NEW_VIEWPORT 0x1 @@ -57,16 +58,20 @@ struct tgsi_sampler; struct vertex_info; - +struct pipe_context; +struct llvmpipe_context; struct lp_fragment_shader; struct lp_fragment_shader_variant_key { + enum pipe_format zsbuf_format; struct pipe_depth_state depth; struct pipe_alpha_state alpha; struct pipe_blend_state blend; + + struct lp_sampler_static_state sampler[PIPE_MAX_SAMPLERS]; }; diff --git a/src/gallium/drivers/llvmpipe/lp_state_blend.c b/src/gallium/drivers/llvmpipe/lp_state_blend.c index 3f03bd0057..b2e75d3b14 100644 --- a/src/gallium/drivers/llvmpipe/lp_state_blend.c +++ b/src/gallium/drivers/llvmpipe/lp_state_blend.c @@ -76,7 +76,7 @@ void llvmpipe_set_blend_color( struct pipe_context *pipe, for (i = 0; i < 4; ++i) { uint8_t c = float_to_ubyte(blend_color->color[i]); for (j = 0; j < 16; ++j) - llvmpipe->jit_context.blend_color[i*4 + j] = c; + llvmpipe->jit_context.blend_color[i*16 + j] = c; } } diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c index 6fbb057937..c753b183c0 100644 --- a/src/gallium/drivers/llvmpipe/lp_state_derived.c +++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c @@ -65,26 +65,19 @@ llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe) if (vinfo->num_attribs == 0) { /* compute vertex layout now */ const struct lp_fragment_shader *lpfs = llvmpipe->fs; - const enum interp_mode colorInterp - = llvmpipe->rasterizer->flatshade ? INTERP_CONSTANT : INTERP_LINEAR; + struct vertex_info *vinfo_vbuf = &llvmpipe->vertex_info_vbuf; + const uint num = draw_num_vs_outputs(llvmpipe->draw); uint i; - if (llvmpipe->vbuf) { - /* if using the post-transform vertex buffer, tell draw_vbuf to - * simply emit the whole post-xform vertex as-is: - */ - struct vertex_info *vinfo_vbuf = &llvmpipe->vertex_info_vbuf; - const uint num = draw_num_vs_outputs(llvmpipe->draw); - uint i; - - /* No longer any need to try and emit draw vertex_header info. - */ - vinfo_vbuf->num_attribs = 0; - for (i = 0; i < num; i++) { - draw_emit_vertex_attr(vinfo_vbuf, EMIT_4F, INTERP_PERSPECTIVE, i); - } - draw_compute_vertex_size(vinfo_vbuf); + /* Tell draw_vbuf to simply emit the whole post-xform vertex + * as-is. No longer any need to try and emit draw vertex_header + * info. + */ + vinfo_vbuf->num_attribs = 0; + for (i = 0; i < num; i++) { + draw_emit_vertex_attr(vinfo_vbuf, EMIT_4F, INTERP_PERSPECTIVE, i); } + draw_compute_vertex_size(vinfo_vbuf); /* * Loop over fragment shader inputs, searching for the matching output @@ -93,35 +86,40 @@ llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe) vinfo->num_attribs = 0; for (i = 0; i < lpfs->info.num_inputs; i++) { int src; - switch (lpfs->info.input_semantic_name[i]) { - case TGSI_SEMANTIC_POSITION: - src = draw_find_vs_output(llvmpipe->draw, - TGSI_SEMANTIC_POSITION, 0); - draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_POS, src); - break; + enum interp_mode interp; - case TGSI_SEMANTIC_COLOR: - src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_COLOR, - lpfs->info.input_semantic_index[i]); - draw_emit_vertex_attr(vinfo, EMIT_4F, colorInterp, src); + switch (lpfs->info.input_interpolate[i]) { + case TGSI_INTERPOLATE_CONSTANT: + interp = INTERP_CONSTANT; break; - - case TGSI_SEMANTIC_FOG: - src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_FOG, 0); - draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src); + case TGSI_INTERPOLATE_LINEAR: + interp = INTERP_LINEAR; break; - - case TGSI_SEMANTIC_GENERIC: - case TGSI_SEMANTIC_FACE: - /* this includes texcoords and varying vars */ - src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_GENERIC, - lpfs->info.input_semantic_index[i]); - draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src); + case TGSI_INTERPOLATE_PERSPECTIVE: + interp = INTERP_PERSPECTIVE; break; - default: assert(0); + interp = INTERP_LINEAR; } + + switch (lpfs->info.input_semantic_name[i]) { + case TGSI_SEMANTIC_POSITION: + interp = INTERP_POS; + break; + + case TGSI_SEMANTIC_COLOR: + if (llvmpipe->rasterizer->flatshade) { + interp = INTERP_CONSTANT; + } + break; + } + + /* this includes texcoords and varying vars */ + src = draw_find_vs_output(llvmpipe->draw, + lpfs->info.input_semantic_name[i], + lpfs->info.input_semantic_index[i]); + draw_emit_vertex_attr(vinfo, EMIT_4F, interp, src); } llvmpipe->psize_slot = draw_find_vs_output(llvmpipe->draw, @@ -250,7 +248,9 @@ void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe ) if (llvmpipe->dirty & (LP_NEW_FS | LP_NEW_BLEND | - LP_NEW_DEPTH_STENCIL_ALPHA)) + LP_NEW_DEPTH_STENCIL_ALPHA | + LP_NEW_SAMPLER | + LP_NEW_TEXTURE)) llvmpipe_update_fs( llvmpipe ); diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c index 94170bd716..2e9aa9fffe 100644 --- a/src/gallium/drivers/llvmpipe/lp_state_fs.c +++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c @@ -83,8 +83,10 @@ #include "lp_bld_debug.h" #include "lp_screen.h" #include "lp_context.h" +#include "lp_buffer.h" #include "lp_state.h" #include "lp_quad.h" +#include "lp_tex_sample.h" static const unsigned char quad_offset_x[4] = {0, 1, 0, 1}; @@ -130,21 +132,20 @@ generate_pos0(LLVMBuilderRef builder, * Generate the depth test. */ static void -generate_depth(struct llvmpipe_context *lp, - LLVMBuilderRef builder, - const struct pipe_depth_state *state, - union lp_type src_type, +generate_depth(LLVMBuilderRef builder, + const struct lp_fragment_shader_variant_key *key, + struct lp_type src_type, struct lp_build_mask_context *mask, LLVMValueRef src, LLVMValueRef dst_ptr) { const struct util_format_description *format_desc; - union lp_type dst_type; + struct lp_type dst_type; - if(!lp->framebuffer.zsbuf) + if(!key->depth.enabled) return; - format_desc = util_format_description(lp->framebuffer.zsbuf->format); + format_desc = util_format_description(key->zsbuf_format); assert(format_desc); /* Pick the depth type. */ @@ -164,7 +165,7 @@ generate_depth(struct llvmpipe_context *lp, #endif lp_build_depth_test(builder, - state, + &key->depth, dst_type, format_desc, mask, @@ -173,107 +174,6 @@ generate_depth(struct llvmpipe_context *lp, } -struct build_fetch_texel_context -{ - LLVMValueRef context_ptr; - - LLVMValueRef samplers_ptr; - - /** Coords/texels store */ - LLVMValueRef store_ptr; -}; - - -void PIPE_CDECL -lp_fetch_texel_soa( struct tgsi_sampler **samplers, - uint32_t unit, - float *store ) -{ - struct tgsi_sampler *sampler = samplers[unit]; - -#if 0 - uint j; - - debug_printf("%s sampler: %p (%p) store: %p\n", - __FUNCTION__, - sampler, *sampler, - store ); - - debug_printf("lodbias %f\n", store[12]); - - for (j = 0; j < 4; j++) - debug_printf("sample %d texcoord %f %f\n", - j, - store[0+j], - store[4+j]); -#endif - - { - float rgba[NUM_CHANNELS][QUAD_SIZE]; - sampler->get_samples(sampler, - &store[0], - &store[4], - &store[8], - 0.0f, /*store[12], lodbias */ - rgba); - memcpy(store, rgba, sizeof rgba); - } - -#if 0 - for (j = 0; j < 4; j++) - debug_printf("sample %d result %f %f %f %f\n", - j, - store[0+j], - store[4+j], - store[8+j], - store[12+j]); -#endif -} - - -static void -emit_fetch_texel( LLVMBuilderRef builder, - void *context, - unsigned unit, - unsigned num_coords, - const LLVMValueRef *coords, - LLVMValueRef lodbias, - LLVMValueRef *texel) -{ - struct build_fetch_texel_context *bld = context; - LLVMTypeRef vec_type = LLVMTypeOf(coords[0]); - LLVMValueRef args[3]; - unsigned i; - - if(!bld->samplers_ptr) - bld->samplers_ptr = lp_jit_context_samplers(builder, bld->context_ptr); - - if(!bld->store_ptr) - bld->store_ptr = LLVMBuildArrayAlloca(builder, - vec_type, - LLVMConstInt(LLVMInt32Type(), 4, 0), - "texel_store"); - - for (i = 0; i < num_coords; i++) { - LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); - LLVMValueRef coord_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, ""); - LLVMBuildStore(builder, coords[i], coord_ptr); - } - - args[0] = bld->samplers_ptr; - args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0); - args[2] = bld->store_ptr; - - lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3); - - for (i = 0; i < NUM_CHANNELS; ++i) { - LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); - LLVMValueRef texel_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, ""); - texel[i] = LLVMBuildLoad(builder, texel_ptr, ""); - } -} - - /** * Generate the fragment shader, depth/stencil test, and alpha tests. */ @@ -282,11 +182,11 @@ generate_fs(struct llvmpipe_context *lp, struct lp_fragment_shader *shader, const struct lp_fragment_shader_variant_key *key, LLVMBuilderRef builder, - union lp_type type, + struct lp_type type, LLVMValueRef context_ptr, unsigned i, const struct lp_build_interp_soa_context *interp, - struct build_fetch_texel_context *sampler, + struct lp_build_sampler_soa *sampler, LLVMValueRef *pmask, LLVMValueRef *color, LLVMValueRef depth_ptr) @@ -298,6 +198,7 @@ generate_fs(struct llvmpipe_context *lp, LLVMValueRef consts_ptr; LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS]; LLVMValueRef z = interp->pos[2]; + struct lp_build_flow_context *flow; struct lp_build_mask_context mask; boolean early_depth_test; unsigned attrib; @@ -309,25 +210,35 @@ generate_fs(struct llvmpipe_context *lp, consts_ptr = lp_jit_context_constants(builder, context_ptr); - lp_build_mask_begin(&mask, builder, type, *pmask); + flow = lp_build_flow_create(builder); + + memset(outputs, 0, sizeof outputs); + + lp_build_flow_scope_begin(flow); + + /* Declare the color and z variables */ + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + color[chan] = LLVMGetUndef(vec_type); + lp_build_flow_scope_declare(flow, &color[chan]); + } + lp_build_flow_scope_declare(flow, &z); + + lp_build_mask_begin(&mask, flow, type, *pmask); early_depth_test = - lp->depth_stencil->depth.enabled && - lp->framebuffer.zsbuf && - !lp->depth_stencil->alpha.enabled && - !lp->fs->info.uses_kill && - !lp->fs->info.writes_z; + key->depth.enabled && + !key->alpha.enabled && + !shader->info.uses_kill && + !shader->info.writes_z; if(early_depth_test) - generate_depth(lp, builder, &key->depth, + generate_depth(builder, key, type, &mask, z, depth_ptr); - memset(outputs, 0, sizeof outputs); - lp_build_tgsi_soa(builder, tokens, type, &mask, consts_ptr, interp->pos, interp->inputs, - outputs, emit_fetch_texel, sampler); + outputs, sampler); for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) { for(chan = 0; chan < NUM_CHANNELS; ++chan) { @@ -368,12 +279,16 @@ generate_fs(struct llvmpipe_context *lp, } if(!early_depth_test) - generate_depth(lp, builder, &key->depth, + generate_depth(builder, key, type, &mask, z, depth_ptr); lp_build_mask_end(&mask); + lp_build_flow_scope_end(flow); + + lp_build_flow_destroy(flow); + *pmask = mask.value; } @@ -385,13 +300,15 @@ generate_fs(struct llvmpipe_context *lp, static void generate_blend(const struct pipe_blend_state *blend, LLVMBuilderRef builder, - union lp_type type, + struct lp_type type, LLVMValueRef context_ptr, LLVMValueRef mask, LLVMValueRef *src, LLVMValueRef dst_ptr) { struct lp_build_context bld; + struct lp_build_flow_context *flow; + struct lp_build_mask_context mask_ctx; LLVMTypeRef vec_type; LLVMTypeRef int_vec_type; LLVMValueRef const_ptr; @@ -400,11 +317,14 @@ generate_blend(const struct pipe_blend_state *blend, LLVMValueRef res[4]; unsigned chan; + lp_build_context_init(&bld, builder, type); + + flow = lp_build_flow_create(builder); + lp_build_mask_begin(&mask_ctx, flow, type, mask); + vec_type = lp_build_vec_type(type); int_vec_type = lp_build_int_vec_type(type); - lp_build_context_init(&bld, builder, type); - const_ptr = lp_jit_context_blend_color(builder, context_ptr); const_ptr = LLVMBuildBitCast(builder, const_ptr, LLVMPointerType(vec_type, 0), ""); @@ -422,11 +342,16 @@ generate_blend(const struct pipe_blend_state *blend, lp_build_blend_soa(builder, blend, type, src, dst, con, res); for(chan = 0; chan < 4; ++chan) { - LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); - lp_build_name(res[chan], "res.%c", "rgba"[chan]); - res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); - LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); + if(blend->colormask & (1 << chan)) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); + lp_build_name(res[chan], "res.%c", "rgba"[chan]); + res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); + LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); + } } + + lp_build_mask_end(&mask_ctx); + lp_build_flow_destroy(flow); } @@ -440,8 +365,8 @@ generate_fragment(struct llvmpipe_context *lp, { struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen); struct lp_fragment_shader_variant *variant; - union lp_type fs_type; - union lp_type blend_type; + struct lp_type fs_type; + struct lp_type blend_type; LLVMTypeRef fs_elem_type; LLVMTypeRef fs_vec_type; LLVMTypeRef fs_int_vec_type; @@ -462,7 +387,7 @@ generate_fragment(struct llvmpipe_context *lp, LLVMBuilderRef builder; LLVMValueRef x0; LLVMValueRef y0; - struct build_fetch_texel_context sampler; + struct lp_build_sampler_soa *sampler; struct lp_build_interp_soa_context interp; LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH]; LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH]; @@ -475,9 +400,9 @@ generate_fragment(struct llvmpipe_context *lp, #ifdef DEBUG tgsi_dump(shader->base.tokens, 0); if(key->depth.enabled) { + debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format)); debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE)); debug_printf("depth.writemask = %u\n", key->depth.writemask); - debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count); } if(key->alpha.enabled) { debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE)); @@ -495,6 +420,34 @@ generate_fragment(struct llvmpipe_context *lp, debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE)); } debug_printf("blend.colormask = 0x%x\n", key->blend.colormask); + for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) { + if(key->sampler[i].format) { + debug_printf("sampler[%u] = \n", i); + debug_printf(" .format = %s\n", + pf_name(key->sampler[i].format)); + debug_printf(" .target = %s\n", + debug_dump_tex_target(key->sampler[i].target, TRUE)); + debug_printf(" .pot = %u %u %u\n", + key->sampler[i].pot_width, + key->sampler[i].pot_height, + key->sampler[i].pot_depth); + debug_printf(" .wrap = %s %s %s\n", + debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE), + debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE), + debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE)); + debug_printf(" .min_img_filter = %s\n", + debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE)); + debug_printf(" .min_mip_filter = %s\n", + debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE)); + debug_printf(" .mag_img_filter = %s\n", + debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE)); + if(key->sampler[i].compare_mode) + debug_printf(" .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE)); + debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords); + debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter); + } + } + #endif variant = CALLOC_STRUCT(lp_fragment_shader_variant); @@ -507,7 +460,7 @@ generate_fragment(struct llvmpipe_context *lp, /* TODO: actually pick these based on the fs and color buffer * characteristics. */ - fs_type.value = 0; + memset(&fs_type, 0, sizeof fs_type); fs_type.floating = TRUE; /* floating point values */ fs_type.sign = TRUE; /* values are signed */ fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ @@ -515,7 +468,7 @@ generate_fragment(struct llvmpipe_context *lp, fs_type.length = 4; /* 4 element per vector */ num_fs = 4; - blend_type.value = 0; + memset(&blend_type, 0, sizeof blend_type); blend_type.floating = FALSE; /* values are integers */ blend_type.sign = FALSE; /* values are unsigned */ blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */ @@ -586,8 +539,13 @@ generate_fragment(struct llvmpipe_context *lp, a0_ptr, dadx_ptr, dady_ptr, x0, y0, 2, 0); - memset(&sampler, 0, sizeof sampler); - sampler.context_ptr = context_ptr; +#if 0 + /* C texture sampling */ + sampler = lp_c_sampler_soa_create(context_ptr); +#else + /* code generated texture sampling */ + sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr); +#endif for(i = 0; i < num_fs; ++i) { LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); @@ -606,7 +564,7 @@ generate_fragment(struct llvmpipe_context *lp, context_ptr, i, &interp, - &sampler, + sampler, &fs_mask[i], out_color, depth_ptr_i); @@ -615,6 +573,8 @@ generate_fragment(struct llvmpipe_context *lp, fs_out_color[chan][i] = out_color[chan]; } + sampler->destroy(sampler); + /* * Convert the fs's output color and mask to fit to the blending type. */ @@ -651,6 +611,11 @@ generate_fragment(struct llvmpipe_context *lp, * Translate the LLVM IR into machine code. */ + if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) { + LLVMDumpValue(variant->function); + abort(); + } + LLVMRunFunctionPassManager(screen->pass, variant->function); #ifdef DEBUG @@ -658,11 +623,6 @@ generate_fragment(struct llvmpipe_context *lp, debug_printf("\n"); #endif - if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) { - LLVMDumpValue(variant->function); - abort(); - } - variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function); #ifdef DEBUG @@ -741,16 +701,29 @@ llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs) void llvmpipe_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index, - const struct pipe_constant_buffer *buf) + const struct pipe_constant_buffer *constants) { struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + struct pipe_buffer *buffer = constants ? constants->buffer : NULL; + unsigned size = buffer ? buffer->size : 0; + const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL; assert(shader < PIPE_SHADER_TYPES); assert(index == 0); + if(shader == PIPE_SHADER_VERTEX) + draw_flush(llvmpipe->draw); + /* note: reference counting */ - pipe_buffer_reference(&llvmpipe->constants[shader].buffer, - buf ? buf->buffer : NULL); + pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer); + + if(shader == PIPE_SHADER_FRAGMENT) { + llvmpipe->jit_context.constants = data; + } + + if(shader == PIPE_SHADER_VERTEX) { + draw_set_mapped_constant_buffer(llvmpipe->draw, data, size); + } llvmpipe->dirty |= LP_NEW_CONSTANTS; } @@ -765,18 +738,45 @@ llvmpipe_set_constant_buffer(struct pipe_context *pipe, */ static void make_variant_key(struct llvmpipe_context *lp, + struct lp_fragment_shader *shader, struct lp_fragment_shader_variant_key *key) { + unsigned i; + memset(key, 0, sizeof *key); - memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); + if(lp->framebuffer.zsbuf && + lp->depth_stencil->depth.enabled) { + key->zsbuf_format = lp->framebuffer.zsbuf->format; + memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); + } key->alpha.enabled = lp->depth_stencil->alpha.enabled; if(key->alpha.enabled) key->alpha.func = lp->depth_stencil->alpha.func; /* alpha.ref_value is passed in jit_context */ - memcpy(&key->blend, lp->blend, sizeof key->blend); + if(lp->framebuffer.cbufs[0]) { + const struct util_format_description *format_desc; + unsigned chan; + + memcpy(&key->blend, lp->blend, sizeof key->blend); + + format_desc = util_format_description(lp->framebuffer.cbufs[0]->format); + assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB || + format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB); + + /* mask out color channels not present in the color buffer */ + for(chan = 0; chan < 4; ++chan) { + enum util_format_swizzle swizzle = format_desc->swizzle[chan]; + if(swizzle > 4) + key->blend.colormask &= ~(1 << chan); + } + } + + for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) + if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) + lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]); } @@ -787,7 +787,7 @@ llvmpipe_update_fs(struct llvmpipe_context *lp) struct lp_fragment_shader_variant_key key; struct lp_fragment_shader_variant *variant; - make_variant_key(lp, &key); + make_variant_key(lp, shader, &key); variant = shader->variants; while(variant) { diff --git a/src/gallium/drivers/llvmpipe/lp_state_sampler.c b/src/gallium/drivers/llvmpipe/lp_state_sampler.c index 4fef541b1e..c69d90c723 100644 --- a/src/gallium/drivers/llvmpipe/lp_state_sampler.c +++ b/src/gallium/drivers/llvmpipe/lp_state_sampler.c @@ -98,6 +98,16 @@ llvmpipe_set_sampler_textures(struct pipe_context *pipe, pipe_texture_reference(&llvmpipe->texture[i], tex); lp_tex_tile_cache_set_texture(llvmpipe->tex_cache[i], tex); + + if(tex) { + struct llvmpipe_texture *lp_tex = llvmpipe_texture(tex); + struct lp_jit_texture *jit_tex = &llvmpipe->jit_context.textures[i]; + jit_tex->width = tex->width[0]; + jit_tex->height = tex->height[0]; + jit_tex->stride = lp_tex->stride[0]; + if(!lp_tex->dt) + jit_tex->data = lp_tex->data; + } } llvmpipe->num_textures = num; diff --git a/src/gallium/drivers/llvmpipe/lp_state_surface.c b/src/gallium/drivers/llvmpipe/lp_state_surface.c index 177a26b7b1..c06ce8b75c 100644 --- a/src/gallium/drivers/llvmpipe/lp_state_surface.c +++ b/src/gallium/drivers/llvmpipe/lp_state_surface.c @@ -53,10 +53,11 @@ llvmpipe_set_framebuffer_state(struct pipe_context *pipe, /* check if changing cbuf */ if (lp->framebuffer.cbufs[i] != fb->cbufs[i]) { /* flush old */ + lp_tile_cache_map_transfers(lp->cbuf_cache[i]); lp_flush_tile_cache(lp->cbuf_cache[i]); /* assign new */ - lp->framebuffer.cbufs[i] = fb->cbufs[i]; + pipe_surface_reference(&lp->framebuffer.cbufs[i], fb->cbufs[i]); /* update cache */ lp_tile_cache_set_surface(lp->cbuf_cache[i], fb->cbufs[i]); @@ -81,7 +82,7 @@ llvmpipe_set_framebuffer_state(struct pipe_context *pipe, } /* assign new */ - lp->framebuffer.zsbuf = fb->zsbuf; + pipe_surface_reference(&lp->framebuffer.zsbuf, fb->zsbuf); /* Tell draw module how deep the Z/depth buffer is */ if (lp->framebuffer.zsbuf) { diff --git a/src/gallium/drivers/llvmpipe/lp_test.h b/src/gallium/drivers/llvmpipe/lp_test.h index 69aaae26e0..39d80726e6 100644 --- a/src/gallium/drivers/llvmpipe/lp_test.h +++ b/src/gallium/drivers/llvmpipe/lp_test.h @@ -56,6 +56,9 @@ #include "lp_bld_type.h" +#define LP_TEST_NUM_SAMPLES 32 + + void write_tsv_header(FILE *fp); @@ -68,17 +71,28 @@ boolean test_all(unsigned verbose, FILE *fp); +#if defined(PIPE_CC_MSVC) + +unsigned __int64 __rdtsc(); +#pragma intrinsic(__rdtsc) +#define rdtsc() __rdtsc() + +#elif defined(PIPE_CC_GCC) && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)) + static INLINE uint64_t rdtsc(void) { -#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) uint32_t hi, lo; __asm__ __volatile__ ("rdtsc" : "=a"(lo), "=d"(hi)); return ((uint64_t)lo) | (((uint64_t)hi) << 32); +} + #else - return 0; + +#define rdtsc() 0 + #endif -} + float @@ -86,43 +100,43 @@ random_float(void); void -dump_type(FILE *fp, union lp_type type); +dump_type(FILE *fp, struct lp_type type); double -read_elem(union lp_type type, const void *src, unsigned index); +read_elem(struct lp_type type, const void *src, unsigned index); void -write_elem(union lp_type type, void *dst, unsigned index, double src); +write_elem(struct lp_type type, void *dst, unsigned index, double src); void -random_elem(union lp_type type, void *dst, unsigned index); +random_elem(struct lp_type type, void *dst, unsigned index); void -read_vec(union lp_type type, const void *src, double *dst); +read_vec(struct lp_type type, const void *src, double *dst); void -write_vec(union lp_type type, void *dst, const double *src); +write_vec(struct lp_type type, void *dst, const double *src); void -random_vec(union lp_type type, void *dst); +random_vec(struct lp_type type, void *dst); boolean -compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps); +compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps); boolean -compare_vec(union lp_type type, const void *res, const void *ref); +compare_vec(struct lp_type type, const void *res, const void *ref); void -dump_vec(FILE *fp, union lp_type type, const void *src); +dump_vec(FILE *fp, struct lp_type type, const void *src); #endif /* !LP_TEST_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c index 8dfad468e3..29fff91981 100644 --- a/src/gallium/drivers/llvmpipe/lp_test_blend.c +++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c @@ -80,7 +80,7 @@ static void write_tsv_row(FILE *fp, const struct pipe_blend_state *blend, enum vector_mode mode, - union lp_type type, + struct lp_type type, double cycles, boolean success) { @@ -125,7 +125,7 @@ static void dump_blend_type(FILE *fp, const struct pipe_blend_state *blend, enum vector_mode mode, - union lp_type type) + struct lp_type type) { fprintf(fp, "%s", mode ? "soa" : "aos"); @@ -153,7 +153,7 @@ static LLVMValueRef add_blend_test(LLVMModuleRef module, const struct pipe_blend_state *blend, enum vector_mode mode, - union lp_type type) + struct lp_type type) { LLVMTypeRef ret_type; LLVMTypeRef vec_type; @@ -462,12 +462,13 @@ compute_blend_ref(const struct pipe_blend_state *blend, } +ALIGN_STACK static boolean test_one(unsigned verbose, FILE *fp, const struct pipe_blend_state *blend, enum vector_mode mode, - union lp_type type) + struct lp_type type) { LLVMModuleRef module = NULL; LLVMValueRef func = NULL; @@ -477,8 +478,8 @@ test_one(unsigned verbose, char *error = NULL; blend_test_ptr_t blend_test_ptr; boolean success; - const unsigned n = 32; - int64_t cycles[n]; + const unsigned n = LP_TEST_NUM_SAMPLES; + int64_t cycles[LP_TEST_NUM_SAMPLES]; double cycles_avg = 0.0; unsigned i, j; @@ -530,11 +531,11 @@ test_one(unsigned verbose, success = TRUE; for(i = 0; i < n && success; ++i) { if(mode == AoS) { - uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t con[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t res[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + ALIGN16_ATTRIB uint8_t src[LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t dst[LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t con[LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t res[LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t ref[LP_NATIVE_VECTOR_WIDTH/8]; int64_t start_counter = 0; int64_t end_counter = 0; @@ -595,11 +596,11 @@ test_one(unsigned verbose, if(mode == SoA) { const unsigned stride = type.length*type.width/8; - uint8_t src[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t dst[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t con[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t res[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; - uint8_t ref[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + ALIGN16_ATTRIB uint8_t src[4*LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t dst[4*LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t con[4*LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t res[4*LP_NATIVE_VECTOR_WIDTH/8]; + ALIGN16_ATTRIB uint8_t ref[4*LP_NATIVE_VECTOR_WIDTH/8]; int64_t start_counter = 0; int64_t end_counter = 0; boolean mismatch; @@ -765,10 +766,10 @@ blend_funcs[] = { }; -const union lp_type blend_types[] = { +const struct lp_type blend_types[] = { /* float, fixed, sign, norm, width, len */ - {{ TRUE, FALSE, FALSE, TRUE, 32, 4 }}, /* f32 x 4 */ - {{ FALSE, FALSE, FALSE, TRUE, 8, 16 }}, /* u8n x 16 */ + { TRUE, FALSE, FALSE, TRUE, 32, 4 }, /* f32 x 4 */ + { FALSE, FALSE, FALSE, TRUE, 8, 16 }, /* u8n x 16 */ }; @@ -788,7 +789,7 @@ test_all(unsigned verbose, FILE *fp) const unsigned *alpha_dst_factor; struct pipe_blend_state blend; enum vector_mode mode; - const union lp_type *type; + const struct lp_type *type; bool success = TRUE; for(rgb_func = blend_funcs; rgb_func < &blend_funcs[num_funcs]; ++rgb_func) { @@ -841,27 +842,27 @@ test_some(unsigned verbose, FILE *fp, unsigned long n) const unsigned *alpha_dst_factor; struct pipe_blend_state blend; enum vector_mode mode; - const union lp_type *type; + const struct lp_type *type; unsigned long i; bool success = TRUE; for(i = 0; i < n; ++i) { - rgb_func = &blend_funcs[random() % num_funcs]; - alpha_func = &blend_funcs[random() % num_funcs]; - rgb_src_factor = &blend_factors[random() % num_factors]; - alpha_src_factor = &blend_factors[random() % num_factors]; + rgb_func = &blend_funcs[rand() % num_funcs]; + alpha_func = &blend_funcs[rand() % num_funcs]; + rgb_src_factor = &blend_factors[rand() % num_factors]; + alpha_src_factor = &blend_factors[rand() % num_factors]; do { - rgb_dst_factor = &blend_factors[random() % num_factors]; + rgb_dst_factor = &blend_factors[rand() % num_factors]; } while(*rgb_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE); do { - alpha_dst_factor = &blend_factors[random() % num_factors]; + alpha_dst_factor = &blend_factors[rand() % num_factors]; } while(*alpha_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE); - mode = random() & 1; + mode = rand() & 1; - type = &blend_types[random() % num_types]; + type = &blend_types[rand() % num_types]; memset(&blend, 0, sizeof blend); blend.blend_enable = 1; diff --git a/src/gallium/drivers/llvmpipe/lp_test_conv.c b/src/gallium/drivers/llvmpipe/lp_test_conv.c index e6489834af..968c7a2d4a 100644 --- a/src/gallium/drivers/llvmpipe/lp_test_conv.c +++ b/src/gallium/drivers/llvmpipe/lp_test_conv.c @@ -59,8 +59,8 @@ write_tsv_header(FILE *fp) static void write_tsv_row(FILE *fp, - union lp_type src_type, - union lp_type dst_type, + struct lp_type src_type, + struct lp_type dst_type, double cycles, boolean success) { @@ -80,8 +80,8 @@ write_tsv_row(FILE *fp, static void dump_conv_types(FILE *fp, - union lp_type src_type, - union lp_type dst_type) + struct lp_type src_type, + struct lp_type dst_type) { fprintf(fp, "src_type="); dump_type(fp, src_type); @@ -96,8 +96,8 @@ dump_conv_types(FILE *fp, static LLVMValueRef add_conv_test(LLVMModuleRef module, - union lp_type src_type, unsigned num_srcs, - union lp_type dst_type, unsigned num_dsts) + struct lp_type src_type, unsigned num_srcs, + struct lp_type dst_type, unsigned num_dsts) { LLVMTypeRef args[2]; LLVMValueRef func; @@ -142,11 +142,12 @@ add_conv_test(LLVMModuleRef module, } +ALIGN_STACK static boolean test_one(unsigned verbose, FILE *fp, - union lp_type src_type, - union lp_type dst_type) + struct lp_type src_type, + struct lp_type dst_type) { LLVMModuleRef module = NULL; LLVMValueRef func = NULL; @@ -156,8 +157,8 @@ test_one(unsigned verbose, char *error = NULL; conv_test_ptr_t conv_test_ptr; boolean success; - const unsigned n = 32; - int64_t cycles[n]; + const unsigned n = LP_TEST_NUM_SAMPLES; + int64_t cycles[LP_TEST_NUM_SAMPLES]; double cycles_avg = 0.0; unsigned num_srcs; unsigned num_dsts; @@ -229,8 +230,8 @@ test_one(unsigned verbose, for(i = 0; i < n && success; ++i) { unsigned src_stride = src_type.length*src_type.width/8; unsigned dst_stride = dst_type.length*dst_type.width/8; - uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; - uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + ALIGN16_ATTRIB uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + ALIGN16_ATTRIB uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; double fref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; int64_t start_counter = 0; @@ -343,35 +344,35 @@ test_one(unsigned verbose, } -const union lp_type conv_types[] = { +const struct lp_type conv_types[] = { /* float, fixed, sign, norm, width, len */ - {{ TRUE, FALSE, TRUE, TRUE, 32, 4 }}, - {{ TRUE, FALSE, TRUE, FALSE, 32, 4 }}, - {{ TRUE, FALSE, FALSE, TRUE, 32, 4 }}, - {{ TRUE, FALSE, FALSE, FALSE, 32, 4 }}, + { TRUE, FALSE, TRUE, TRUE, 32, 4 }, + { TRUE, FALSE, TRUE, FALSE, 32, 4 }, + { TRUE, FALSE, FALSE, TRUE, 32, 4 }, + { TRUE, FALSE, FALSE, FALSE, 32, 4 }, /* TODO: test fixed formats too */ - {{ FALSE, FALSE, TRUE, TRUE, 16, 8 }}, - {{ FALSE, FALSE, TRUE, FALSE, 16, 8 }}, - {{ FALSE, FALSE, FALSE, TRUE, 16, 8 }}, - {{ FALSE, FALSE, FALSE, FALSE, 16, 8 }}, - - {{ FALSE, FALSE, TRUE, TRUE, 32, 4 }}, - {{ FALSE, FALSE, TRUE, FALSE, 32, 4 }}, - {{ FALSE, FALSE, FALSE, TRUE, 32, 4 }}, - {{ FALSE, FALSE, FALSE, FALSE, 32, 4 }}, - - {{ FALSE, FALSE, TRUE, TRUE, 16, 8 }}, - {{ FALSE, FALSE, TRUE, FALSE, 16, 8 }}, - {{ FALSE, FALSE, FALSE, TRUE, 16, 8 }}, - {{ FALSE, FALSE, FALSE, FALSE, 16, 8 }}, - - {{ FALSE, FALSE, TRUE, TRUE, 8, 16 }}, - {{ FALSE, FALSE, TRUE, FALSE, 8, 16 }}, - {{ FALSE, FALSE, FALSE, TRUE, 8, 16 }}, - {{ FALSE, FALSE, FALSE, FALSE, 8, 16 }}, + { FALSE, FALSE, TRUE, TRUE, 16, 8 }, + { FALSE, FALSE, TRUE, FALSE, 16, 8 }, + { FALSE, FALSE, FALSE, TRUE, 16, 8 }, + { FALSE, FALSE, FALSE, FALSE, 16, 8 }, + + { FALSE, FALSE, TRUE, TRUE, 32, 4 }, + { FALSE, FALSE, TRUE, FALSE, 32, 4 }, + { FALSE, FALSE, FALSE, TRUE, 32, 4 }, + { FALSE, FALSE, FALSE, FALSE, 32, 4 }, + + { FALSE, FALSE, TRUE, TRUE, 16, 8 }, + { FALSE, FALSE, TRUE, FALSE, 16, 8 }, + { FALSE, FALSE, FALSE, TRUE, 16, 8 }, + { FALSE, FALSE, FALSE, FALSE, 16, 8 }, + + { FALSE, FALSE, TRUE, TRUE, 8, 16 }, + { FALSE, FALSE, TRUE, FALSE, 8, 16 }, + { FALSE, FALSE, FALSE, TRUE, 8, 16 }, + { FALSE, FALSE, FALSE, FALSE, 8, 16 }, }; @@ -381,8 +382,8 @@ const unsigned num_types = sizeof(conv_types)/sizeof(conv_types[0]); boolean test_all(unsigned verbose, FILE *fp) { - const union lp_type *src_type; - const union lp_type *dst_type; + const struct lp_type *src_type; + const struct lp_type *dst_type; bool success = TRUE; for(src_type = conv_types; src_type < &conv_types[num_types]; ++src_type) { @@ -407,16 +408,16 @@ test_all(unsigned verbose, FILE *fp) boolean test_some(unsigned verbose, FILE *fp, unsigned long n) { - const union lp_type *src_type; - const union lp_type *dst_type; + const struct lp_type *src_type; + const struct lp_type *dst_type; unsigned long i; bool success = TRUE; for(i = 0; i < n; ++i) { - src_type = &conv_types[random() % num_types]; + src_type = &conv_types[rand() % num_types]; do { - dst_type = &conv_types[random() % num_types]; + dst_type = &conv_types[rand() % num_types]; } while (src_type == dst_type || src_type->norm != dst_type->norm); if(!test_one(verbose, fp, *src_type, *dst_type)) diff --git a/src/gallium/drivers/llvmpipe/lp_test_format.c b/src/gallium/drivers/llvmpipe/lp_test_format.c index 1d192355ee..23ea9ebbe7 100644 --- a/src/gallium/drivers/llvmpipe/lp_test_format.c +++ b/src/gallium/drivers/llvmpipe/lp_test_format.c @@ -35,10 +35,11 @@ #include <llvm-c/Target.h> #include <llvm-c/Transforms/Scalar.h> +#include "util/u_cpu_detect.h" #include "util/u_format.h" -#include "lp_bld_flow.h" #include "lp_bld_format.h" +#include "lp_test.h" struct pixel_test_case @@ -89,40 +90,63 @@ struct pixel_test_case test_cases[] = }; -typedef void (*load_ptr_t)(const void *, float *); +void +write_tsv_header(FILE *fp) +{ + fprintf(fp, + "result\t" + "format\n"); + + fflush(fp); +} + + +static void +write_tsv_row(FILE *fp, + const struct util_format_description *desc, + boolean success) +{ + fprintf(fp, "%s\t", success ? "pass" : "fail"); + + fprintf(fp, "%s\n", desc->name); + + fflush(fp); +} + + +typedef void (*load_ptr_t)(const uint32_t packed, float *); static LLVMValueRef add_load_rgba_test(LLVMModuleRef module, - enum pipe_format format) + const struct util_format_description *desc) { LLVMTypeRef args[2]; LLVMValueRef func; - LLVMValueRef ptr; + LLVMValueRef packed; LLVMValueRef rgba_ptr; LLVMBasicBlockRef block; LLVMBuilderRef builder; LLVMValueRef rgba; - struct lp_build_loop_state loop; - args[0] = LLVMPointerType(LLVMInt8Type(), 0); + args[0] = LLVMInt32Type(); args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0); func = LLVMAddFunction(module, "load", LLVMFunctionType(LLVMVoidType(), args, 2, 0)); LLVMSetFunctionCallConv(func, LLVMCCallConv); - ptr = LLVMGetParam(func, 0); + packed = LLVMGetParam(func, 0); rgba_ptr = LLVMGetParam(func, 1); block = LLVMAppendBasicBlock(func, "entry"); builder = LLVMCreateBuilder(); LLVMPositionBuilderAtEnd(builder, block); - lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 1, 0), &loop); + if(desc->block.bits < 32) + packed = LLVMBuildTrunc(builder, packed, LLVMIntType(desc->block.bits), ""); - rgba = lp_build_load_rgba(builder, format, ptr); - LLVMBuildStore(builder, rgba, rgba_ptr); + rgba = lp_build_unpack_rgba_aos(builder, desc, packed); - lp_build_loop_end(builder, LLVMConstInt(LLVMInt32Type(), 4, 0), NULL, &loop); + LLVMBuildStore(builder, rgba, rgba_ptr); LLVMBuildRetVoid(builder); @@ -131,27 +155,28 @@ add_load_rgba_test(LLVMModuleRef module, } -typedef void (*store_ptr_t)(void *, const float *); +typedef void (*store_ptr_t)(uint32_t *, const float *); static LLVMValueRef add_store_rgba_test(LLVMModuleRef module, - enum pipe_format format) + const struct util_format_description *desc) { LLVMTypeRef args[2]; LLVMValueRef func; - LLVMValueRef ptr; + LLVMValueRef packed_ptr; LLVMValueRef rgba_ptr; LLVMBasicBlockRef block; LLVMBuilderRef builder; LLVMValueRef rgba; + LLVMValueRef packed; - args[0] = LLVMPointerType(LLVMInt8Type(), 0); + args[0] = LLVMPointerType(LLVMInt32Type(), 0); args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0); func = LLVMAddFunction(module, "store", LLVMFunctionType(LLVMVoidType(), args, 2, 0)); LLVMSetFunctionCallConv(func, LLVMCCallConv); - ptr = LLVMGetParam(func, 0); + packed_ptr = LLVMGetParam(func, 0); rgba_ptr = LLVMGetParam(func, 1); block = LLVMAppendBasicBlock(func, "entry"); @@ -160,7 +185,12 @@ add_store_rgba_test(LLVMModuleRef module, rgba = LLVMBuildLoad(builder, rgba_ptr, ""); - lp_build_store_rgba(builder, format, ptr, rgba); + packed = lp_build_pack_rgba_aos(builder, desc, rgba); + + if(desc->block.bits < 32) + packed = LLVMBuildZExt(builder, packed, LLVMInt32Type(), ""); + + LLVMBuildStore(builder, packed, packed_ptr); LLVMBuildRetVoid(builder); @@ -169,8 +199,9 @@ add_store_rgba_test(LLVMModuleRef module, } +ALIGN_STACK static boolean -test_format(const struct pixel_test_case *test) +test_format(unsigned verbose, FILE *fp, const struct pixel_test_case *test) { LLVMModuleRef module = NULL; LLVMValueRef load = NULL; @@ -192,8 +223,8 @@ test_format(const struct pixel_test_case *test) module = LLVMModuleCreateWithName("test"); - load = add_load_rgba_test(module, test->format); - store = add_store_rgba_test(module, test->format); + load = add_load_rgba_test(module, desc); + store = add_store_rgba_test(module, desc); if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) { LLVMDumpModule(module); @@ -229,7 +260,7 @@ test_format(const struct pixel_test_case *test) memset(unpacked, 0, sizeof unpacked); packed = 0; - load_ptr(&test->packed, unpacked); + load_ptr(test->packed, unpacked); store_ptr(&packed, unpacked); success = TRUE; @@ -255,18 +286,29 @@ test_format(const struct pixel_test_case *test) if(pass) LLVMDisposePassManager(pass); + if(fp) + write_tsv_row(fp, desc, success); + return success; } -int main(int argc, char **argv) +boolean +test_all(unsigned verbose, FILE *fp) { unsigned i; - int ret; + bool success = TRUE; for (i = 0; i < sizeof(test_cases)/sizeof(test_cases[0]); ++i) - if(!test_format(&test_cases[i])) - ret = 1; + if(!test_format(verbose, fp, &test_cases[i])) + success = FALSE; - return ret; + return success; +} + + +boolean +test_some(unsigned verbose, FILE *fp, unsigned long n) +{ + return test_all(verbose, fp); } diff --git a/src/gallium/drivers/llvmpipe/lp_test_main.c b/src/gallium/drivers/llvmpipe/lp_test_main.c index 49213fb4f0..314544aa9a 100644 --- a/src/gallium/drivers/llvmpipe/lp_test_main.c +++ b/src/gallium/drivers/llvmpipe/lp_test_main.c @@ -34,13 +34,28 @@ */ +#include "util/u_cpu_detect.h" + #include "lp_bld_const.h" +#include "lp_bld_misc.h" #include "lp_test.h" +#ifdef PIPE_CC_MSVC +static INLINE double +round(double x) +{ + if (x >= 0.0) + return floor(x + 0.5); + else + return ceil(x - 0.5); +} +#endif + + void dump_type(FILE *fp, - union lp_type type) + struct lp_type type) { fprintf(fp, "%s%s%u%sx%u", type.sign ? (type.floating || type.fixed ? "" : "s") : "u", @@ -52,7 +67,7 @@ dump_type(FILE *fp, double -read_elem(union lp_type type, const void *src, unsigned index) +read_elem(struct lp_type type, const void *src, unsigned index) { double scale = lp_const_scale(type); double value; @@ -115,7 +130,7 @@ read_elem(union lp_type type, const void *src, unsigned index) void -write_elem(union lp_type type, void *dst, unsigned index, double value) +write_elem(struct lp_type type, void *dst, unsigned index, double value) { assert(index < type.length); if(!type.sign && value < 0.0) @@ -184,11 +199,11 @@ write_elem(union lp_type type, void *dst, unsigned index, double value) void -random_elem(union lp_type type, void *dst, unsigned index) +random_elem(struct lp_type type, void *dst, unsigned index) { double value; assert(index < type.length); - value = (double)random()/(double)RAND_MAX; + value = (double)rand()/(double)RAND_MAX; if(!type.norm) { unsigned long long mask; if (type.floating) @@ -199,17 +214,17 @@ random_elem(union lp_type type, void *dst, unsigned index) mask = ((unsigned long long)1 << (type.width - 1)) - 1; else mask = ((unsigned long long)1 << type.width) - 1; - value += (double)(mask & random()); + value += (double)(mask & rand()); } if(!type.sign) - if(random() & 1) + if(rand() & 1) value = -value; write_elem(type, dst, index, value); } void -read_vec(union lp_type type, const void *src, double *dst) +read_vec(struct lp_type type, const void *src, double *dst) { unsigned i; for (i = 0; i < type.length; ++i) @@ -218,7 +233,7 @@ read_vec(union lp_type type, const void *src, double *dst) void -write_vec(union lp_type type, void *dst, const double *src) +write_vec(struct lp_type type, void *dst, const double *src) { unsigned i; for (i = 0; i < type.length; ++i) @@ -229,12 +244,12 @@ write_vec(union lp_type type, void *dst, const double *src) float random_float(void) { - return (float)((double)random()/(double)RAND_MAX); + return (float)((double)rand()/(double)RAND_MAX); } void -random_vec(union lp_type type, void *dst) +random_vec(struct lp_type type, void *dst) { unsigned i; for (i = 0; i < type.length; ++i) @@ -243,7 +258,7 @@ random_vec(union lp_type type, void *dst) boolean -compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps) +compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps) { unsigned i; for (i = 0; i < type.length; ++i) { @@ -259,7 +274,7 @@ compare_vec_with_eps(union lp_type type, const void *res, const void *ref, doubl boolean -compare_vec(union lp_type type, const void *res, const void *ref) +compare_vec(struct lp_type type, const void *res, const void *ref) { double eps = lp_const_eps(type); return compare_vec_with_eps(type, res, ref, eps); @@ -267,7 +282,7 @@ compare_vec(union lp_type type, const void *res, const void *ref) void -dump_vec(FILE *fp, union lp_type type, const void *src) +dump_vec(FILE *fp, struct lp_type type, const void *src) { unsigned i; for (i = 0; i < type.length; ++i) { @@ -365,6 +380,11 @@ int main(int argc, char **argv) n = atoi(argv[i]); } + LLVMLinkInJIT(); + LLVMInitializeNativeTarget(); + + util_cpu_detect(); + if(fp) { /* Warm up the caches */ test_some(0, NULL, 100); diff --git a/src/gallium/drivers/llvmpipe/lp_tex_cache.c b/src/gallium/drivers/llvmpipe/lp_tex_cache.c index 23a94b5b0d..773e848242 100644 --- a/src/gallium/drivers/llvmpipe/lp_tex_cache.c +++ b/src/gallium/drivers/llvmpipe/lp_tex_cache.c @@ -154,7 +154,7 @@ lp_tex_tile_cache_validate_texture(struct llvmpipe_tex_tile_cache *tc) if (lpt->timestamp != tc->timestamp) { /* texture was modified, invalidate all cached tiles */ uint i; - _debug_printf("INV %d %d\n", tc->timestamp, lpt->timestamp); + debug_printf("INV %d %d\n", tc->timestamp, lpt->timestamp); for (i = 0; i < NUM_ENTRIES; i++) { tc->entries[i].addr.bits.invalid = 1; } diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.h b/src/gallium/drivers/llvmpipe/lp_tex_sample.h index 628ec3f1ef..9ad1bde956 100644 --- a/src/gallium/drivers/llvmpipe/lp_tex_sample.h +++ b/src/gallium/drivers/llvmpipe/lp_tex_sample.h @@ -29,10 +29,13 @@ #define LP_TEX_SAMPLE_H +#include <llvm-c/Core.h> + #include "tgsi/tgsi_exec.h" struct llvmpipe_tex_tile_cache; +struct lp_sampler_static_state; /** @@ -75,4 +78,24 @@ lp_get_samples(struct tgsi_sampler *tgsi_sampler, float rgba[NUM_CHANNELS][QUAD_SIZE]); +/** + * Texture sampling code generator that just calls lp_get_samples C function + * for the actual sampling computation. + * + * @param context_ptr LLVM value with the pointer to the struct lp_jit_context. + */ +struct lp_build_sampler_soa * +lp_c_sampler_soa_create(LLVMValueRef context_ptr); + + +/** + * Pure-LLVM texture sampling code generator. + * + * @param context_ptr LLVM value with the pointer to the struct lp_jit_context. + */ +struct lp_build_sampler_soa * +lp_llvm_sampler_soa_create(const struct lp_sampler_static_state *key, + LLVMValueRef context_ptr); + + #endif /* LP_TEX_SAMPLE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.c b/src/gallium/drivers/llvmpipe/lp_tex_sample_c.c index 94eb6dad5a..a1365a045f 100644 --- a/src/gallium/drivers/llvmpipe/lp_tex_sample.c +++ b/src/gallium/drivers/llvmpipe/lp_tex_sample_c.c @@ -1578,3 +1578,136 @@ out: tgsi_sampler->get_samples( tgsi_sampler, s, t, p, lodbias, rgba ); } + +void PIPE_CDECL +lp_fetch_texel_soa( struct tgsi_sampler **samplers, + uint32_t unit, + float *store ) +{ + struct tgsi_sampler *sampler = samplers[unit]; + +#if 0 + uint j; + + debug_printf("%s sampler: %p (%p) store: %p\n", + __FUNCTION__, + sampler, *sampler, + store ); + + debug_printf("lodbias %f\n", store[12]); + + for (j = 0; j < 4; j++) + debug_printf("sample %d texcoord %f %f\n", + j, + store[0+j], + store[4+j]); +#endif + + { + float rgba[NUM_CHANNELS][QUAD_SIZE]; + sampler->get_samples(sampler, + &store[0], + &store[4], + &store[8], + 0.0f, /*store[12], lodbias */ + rgba); + memcpy(store, rgba, sizeof rgba); + } + +#if 0 + for (j = 0; j < 4; j++) + debug_printf("sample %d result %f %f %f %f\n", + j, + store[0+j], + store[4+j], + store[8+j], + store[12+j]); +#endif +} + + +#include "lp_bld_type.h" +#include "lp_bld_intr.h" +#include "lp_bld_tgsi.h" + + +struct lp_c_sampler_soa +{ + struct lp_build_sampler_soa base; + + LLVMValueRef context_ptr; + + LLVMValueRef samplers_ptr; + + /** Coords/texels store */ + LLVMValueRef store_ptr; +}; + + +static void +lp_c_sampler_soa_destroy(struct lp_build_sampler_soa *sampler) +{ + FREE(sampler); +} + + +static void +lp_c_sampler_soa_emit_fetch_texel(struct lp_build_sampler_soa *_sampler, + LLVMBuilderRef builder, + struct lp_type type, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel) +{ + struct lp_c_sampler_soa *sampler = (struct lp_c_sampler_soa *)_sampler; + LLVMTypeRef vec_type = LLVMTypeOf(coords[0]); + LLVMValueRef args[3]; + unsigned i; + + if(!sampler->samplers_ptr) + sampler->samplers_ptr = lp_jit_context_samplers(builder, sampler->context_ptr); + + if(!sampler->store_ptr) + sampler->store_ptr = LLVMBuildArrayAlloca(builder, + vec_type, + LLVMConstInt(LLVMInt32Type(), 4, 0), + "texel_store"); + + for (i = 0; i < num_coords; i++) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef coord_ptr = LLVMBuildGEP(builder, sampler->store_ptr, &index, 1, ""); + LLVMBuildStore(builder, coords[i], coord_ptr); + } + + args[0] = sampler->samplers_ptr; + args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0); + args[2] = sampler->store_ptr; + + lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3); + + for (i = 0; i < NUM_CHANNELS; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef texel_ptr = LLVMBuildGEP(builder, sampler->store_ptr, &index, 1, ""); + texel[i] = LLVMBuildLoad(builder, texel_ptr, ""); + } +} + + +struct lp_build_sampler_soa * +lp_c_sampler_soa_create(LLVMValueRef context_ptr) +{ + struct lp_c_sampler_soa *sampler; + + sampler = CALLOC_STRUCT(lp_c_sampler_soa); + if(!sampler) + return NULL; + + sampler->base.destroy = lp_c_sampler_soa_destroy; + sampler->base.emit_fetch_texel = lp_c_sampler_soa_emit_fetch_texel; + sampler->context_ptr = context_ptr; + + return &sampler->base; +} + diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c b/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c new file mode 100644 index 0000000000..d2a6ae21f5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c @@ -0,0 +1,196 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All rights reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Texture sampling code generation + * + * This file is nothing more than ugly glue between three largely independent + * entities: + * - TGSI -> LLVM translation (i.e., lp_build_tgsi_soa) + * - texture sampling code generation (i.e., lp_build_sample_soa) + * - LLVM pipe driver + * + * All interesting code is in the functions mentioned above. There is really + * nothing to see here. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_defines.h" +#include "pipe/p_shader_tokens.h" +#include "lp_bld_debug.h" +#include "lp_bld_type.h" +#include "lp_bld_intr.h" +#include "lp_bld_sample.h" +#include "lp_bld_tgsi.h" +#include "lp_state.h" +#include "lp_tex_sample.h" + + +/** + * This provides the bridge between the sampler state store in lp_jit_context + * and lp_jit_texture and the sampler code generator. It provides the + * texture layout information required by the texture sampler code generator + * in terms of the state stored in lp_jit_context and lp_jit_texture in runtime. + */ +struct llvmpipe_sampler_dynamic_state +{ + struct lp_sampler_dynamic_state base; + + const struct lp_sampler_static_state *static_state; + + LLVMValueRef context_ptr; +}; + + +/** + * This is the bridge between our sampler and the TGSI translator. + */ +struct lp_llvm_sampler_soa +{ + struct lp_build_sampler_soa base; + + struct llvmpipe_sampler_dynamic_state dynamic_state; +}; + + +/** + * Fetch the specified member of the lp_jit_texture structure. + * + * @sa http://llvm.org/docs/GetElementPtr.html + */ +static LLVMValueRef +lp_llvm_texture_member(struct lp_sampler_dynamic_state *base, + LLVMBuilderRef builder, + unsigned unit, + unsigned member_index, + const char *member_name) +{ + struct llvmpipe_sampler_dynamic_state *state = (struct llvmpipe_sampler_dynamic_state *)base; + LLVMValueRef indices[4]; + LLVMValueRef ptr; + LLVMValueRef res; + + assert(unit < PIPE_MAX_SAMPLERS); + + /* context[0] */ + indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); + /* context[0].textures */ + indices[1] = LLVMConstInt(LLVMInt32Type(), LP_JIT_CONTEXT_TEXTURES_INDEX, 0); + /* context[0].textures[unit] */ + indices[2] = LLVMConstInt(LLVMInt32Type(), unit, 0); + /* context[0].textures[unit].member */ + indices[3] = LLVMConstInt(LLVMInt32Type(), member_index, 0); + + ptr = LLVMBuildGEP(builder, state->context_ptr, indices, Elements(indices), ""); + + res = LLVMBuildLoad(builder, ptr, ""); + + lp_build_name(res, "context.texture%u.%s", unit, member_name); + + return res; +} + + +/** + * Helper macro to instantiate the functions that generate the code to fetch + * the members of lp_jit_texture to fulfill the sampler code generator requests. + * + * This complexity is the price we have to pay to keep the texture sampler code + * generator a reusable module without dependencies to llvmpipe internals. + */ +#define LP_LLVM_TEXTURE_MEMBER(_name, _index) \ + static LLVMValueRef \ + lp_llvm_texture_##_name( struct lp_sampler_dynamic_state *base, \ + LLVMBuilderRef builder, \ + unsigned unit) \ + { \ + return lp_llvm_texture_member(base, builder, unit, _index, #_name ); \ + } + + +LP_LLVM_TEXTURE_MEMBER(width, LP_JIT_TEXTURE_WIDTH) +LP_LLVM_TEXTURE_MEMBER(height, LP_JIT_TEXTURE_HEIGHT) +LP_LLVM_TEXTURE_MEMBER(stride, LP_JIT_TEXTURE_STRIDE) +LP_LLVM_TEXTURE_MEMBER(data_ptr, LP_JIT_TEXTURE_DATA) + + +static void +lp_llvm_sampler_soa_destroy(struct lp_build_sampler_soa *sampler) +{ + FREE(sampler); +} + + +static void +lp_llvm_sampler_soa_emit_fetch_texel(struct lp_build_sampler_soa *base, + LLVMBuilderRef builder, + struct lp_type type, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel) +{ + struct lp_llvm_sampler_soa *sampler = (struct lp_llvm_sampler_soa *)base; + + assert(unit < PIPE_MAX_SAMPLERS); + + lp_build_sample_soa(builder, + &sampler->dynamic_state.static_state[unit], + &sampler->dynamic_state.base, + type, + unit, + num_coords, + coords, + lodbias, + texel); +} + + +struct lp_build_sampler_soa * +lp_llvm_sampler_soa_create(const struct lp_sampler_static_state *static_state, + LLVMValueRef context_ptr) +{ + struct lp_llvm_sampler_soa *sampler; + + sampler = CALLOC_STRUCT(lp_llvm_sampler_soa); + if(!sampler) + return NULL; + + sampler->base.destroy = lp_llvm_sampler_soa_destroy; + sampler->base.emit_fetch_texel = lp_llvm_sampler_soa_emit_fetch_texel; + sampler->dynamic_state.base.width = lp_llvm_texture_width; + sampler->dynamic_state.base.height = lp_llvm_texture_height; + sampler->dynamic_state.base.stride = lp_llvm_texture_stride; + sampler->dynamic_state.base.data_ptr = lp_llvm_texture_data_ptr; + sampler->dynamic_state.static_state = static_state; + sampler->dynamic_state.context_ptr = context_ptr; + + return &sampler->base; +} + diff --git a/src/gallium/drivers/llvmpipe/lp_texture.c b/src/gallium/drivers/llvmpipe/lp_texture.c index 724d437833..a00f2495df 100644 --- a/src/gallium/drivers/llvmpipe/lp_texture.c +++ b/src/gallium/drivers/llvmpipe/lp_texture.c @@ -66,16 +66,24 @@ llvmpipe_texture_layout(struct llvmpipe_screen *screen, pf_get_block(lpt->base.format, &lpt->base.block); for (level = 0; level <= pt->last_level; level++) { + unsigned nblocksx, nblocksy; + pt->width[level] = width; pt->height[level] = height; pt->depth[level] = depth; pt->nblocksx[level] = pf_get_nblocksx(&pt->block, width); - pt->nblocksy[level] = pf_get_nblocksy(&pt->block, height); - lpt->stride[level] = align(pt->nblocksx[level]*pt->block.size, 16); + pt->nblocksy[level] = pf_get_nblocksy(&pt->block, height); + + /* Allocate storage for whole quads. This is particularly important + * for depth surfaces, which are currently stored in a swizzled format. */ + nblocksx = pf_get_nblocksx(&pt->block, align(width, 2)); + nblocksy = pf_get_nblocksy(&pt->block, align(height, 2)); + + lpt->stride[level] = align(nblocksx*pt->block.size, 16); lpt->level_offset[level] = buffer_size; - buffer_size += (pt->nblocksy[level] * + buffer_size += (nblocksy * ((pt->target == PIPE_TEXTURE_CUBE) ? 6 : depth) * lpt->stride[level]); @@ -353,17 +361,9 @@ llvmpipe_transfer_map( struct pipe_screen *_screen, if(lpt->dt) { struct llvmpipe_winsys *winsys = screen->winsys; - unsigned flags = 0; - - if (transfer->usage != PIPE_TRANSFER_READ) { - flags |= PIPE_BUFFER_USAGE_CPU_WRITE; - } - - if (transfer->usage != PIPE_TRANSFER_WRITE) { - flags |= PIPE_BUFFER_USAGE_CPU_READ; - } - map = winsys->displaytarget_map(winsys, lpt->dt, flags); + map = winsys->displaytarget_map(winsys, lpt->dt, + pipe_transfer_buffer_flags(transfer)); if (map == NULL) return NULL; } @@ -373,7 +373,7 @@ llvmpipe_transfer_map( struct pipe_screen *_screen, /* May want to different things here depending on read/write nature * of the map: */ - if (transfer->texture && transfer->usage != PIPE_TRANSFER_READ) + if (transfer->texture && (transfer->usage & PIPE_TRANSFER_WRITE)) { /* Do something to notify sharing contexts of a texture change. * In llvmpipe, that would mean flushing the texture cache. diff --git a/src/gallium/drivers/llvmpipe/lp_tile_cache.c b/src/gallium/drivers/llvmpipe/lp_tile_cache.c index 143afec3d3..ec3e002d62 100644 --- a/src/gallium/drivers/llvmpipe/lp_tile_cache.c +++ b/src/gallium/drivers/llvmpipe/lp_tile_cache.c @@ -44,10 +44,53 @@ #include "lp_tile_cache.h" +#define MAX_WIDTH 4096 +#define MAX_HEIGHT 4096 + + +enum llvmpipe_tile_status +{ + LP_TILE_STATUS_UNDEFINED = 0, + LP_TILE_STATUS_CLEAR = 1, + LP_TILE_STATUS_DEFINED = 2 +}; + + +struct llvmpipe_cached_tile +{ + enum llvmpipe_tile_status status; + + /** color in SOA format */ + uint8_t *color; +}; + + +struct llvmpipe_tile_cache +{ + struct pipe_screen *screen; + struct pipe_surface *surface; /**< the surface we're caching */ + struct pipe_transfer *transfer; + void *transfer_map; + + struct llvmpipe_cached_tile entries[MAX_WIDTH/TILE_SIZE][MAX_HEIGHT/TILE_SIZE]; + + uint8_t clear_color[4]; /**< for color bufs */ + uint clear_val; /**< for z+stencil, or packed color clear value */ + + struct llvmpipe_cached_tile *last_tile; /**< most recently retrieved tile */ +}; + + struct llvmpipe_tile_cache * lp_create_tile_cache( struct pipe_screen *screen ) { struct llvmpipe_tile_cache *tc; + int maxLevels, maxTexSize; + + /* sanity checking: max sure MAX_WIDTH/HEIGHT >= largest texture image */ + maxLevels = screen->get_param(screen, PIPE_CAP_MAX_TEXTURE_2D_LEVELS); + maxTexSize = 1 << (maxLevels - 1); + assert(MAX_WIDTH >= maxTexSize); tc = CALLOC_STRUCT( llvmpipe_tile_cache ); if(!tc) @@ -193,44 +236,41 @@ lp_flush_tile_cache(struct llvmpipe_tile_cache *tc) if(!pt) return; + assert(tc->transfer_map); + /* push the tile to all positions marked as clear */ for (y = 0; y < pt->height; y += TILE_SIZE) { for (x = 0; x < pt->width; x += TILE_SIZE) { struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; - switch(tile->status) { - case LP_TILE_STATUS_UNDEFINED: - break; - - case LP_TILE_STATUS_CLEAR: { - /** - * Actually clear the tiles which were flagged as being in a clear state. - */ - - struct pipe_screen *screen = pt->texture->screen; - unsigned tw = TILE_SIZE; - unsigned th = TILE_SIZE; - void *dst; + if(tile->status != LP_TILE_STATUS_UNDEFINED) { + unsigned w = TILE_SIZE; + unsigned h = TILE_SIZE; + + if (!pipe_clip_tile(x, y, &w, &h, pt)) { + switch(tile->status) { + case LP_TILE_STATUS_CLEAR: + /* Actually clear the tiles which were flagged as being in a + * clear state. */ + util_fill_rect(tc->transfer_map, &pt->block, pt->stride, + x, y, w, h, + tc->clear_val); + break; + + case LP_TILE_STATUS_DEFINED: + lp_tile_write_4ub(pt->format, + tile->color, + tc->transfer_map, pt->stride, + x, y, w, h); + break; + + default: + assert(0); + break; + } + } - if (pipe_clip_tile(x, y, &tw, &th, pt)) - continue; - - dst = screen->transfer_map(screen, pt); - assert(dst); - if(!dst) - continue; - - util_fill_rect(dst, &pt->block, pt->stride, - x, y, tw, th, - tc->clear_val); - - screen->transfer_unmap(screen, pt); - break; - } - - case LP_TILE_STATUS_DEFINED: - lp_put_tile_rgba_soa(pt, x, y, tile->color); - break; + tile->status = LP_TILE_STATUS_UNDEFINED; } } } @@ -248,6 +288,9 @@ lp_get_cached_tile(struct llvmpipe_tile_cache *tc, struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; struct pipe_transfer *pt = tc->transfer; + assert(tc->surface); + assert(tc->transfer); + switch(tile->status) { case LP_TILE_STATUS_CLEAR: /* don't get tile from framebuffer, just clear it */ @@ -255,11 +298,22 @@ lp_get_cached_tile(struct llvmpipe_tile_cache *tc, tile->status = LP_TILE_STATUS_DEFINED; break; - case LP_TILE_STATUS_UNDEFINED: - /* get new tile data from transfer */ - lp_get_tile_rgba_soa(pt, x, y, tile->color); + case LP_TILE_STATUS_UNDEFINED: { + unsigned w = TILE_SIZE; + unsigned h = TILE_SIZE; + + x &= ~(TILE_SIZE - 1); + y &= ~(TILE_SIZE - 1); + + if (!pipe_clip_tile(x, y, &w, &h, tc->transfer)) + lp_tile_read_4ub(pt->format, + tile->color, + tc->transfer_map, tc->transfer->stride, + x, y, w, h); + tile->status = LP_TILE_STATUS_DEFINED; break; + } case LP_TILE_STATUS_DEFINED: /* nothing to do */ diff --git a/src/gallium/drivers/llvmpipe/lp_tile_cache.h b/src/gallium/drivers/llvmpipe/lp_tile_cache.h index 6d8ba5ece7..161bab3799 100644 --- a/src/gallium/drivers/llvmpipe/lp_tile_cache.h +++ b/src/gallium/drivers/llvmpipe/lp_tile_cache.h @@ -33,42 +33,7 @@ #include "lp_tile_soa.h" -enum llvmpipe_tile_status -{ - LP_TILE_STATUS_UNDEFINED = 0, - LP_TILE_STATUS_CLEAR = 1, - LP_TILE_STATUS_DEFINED = 2 -}; - - -struct llvmpipe_cached_tile -{ - enum llvmpipe_tile_status status; - - /** color in SOA format */ - uint8_t *color; -}; - - -/** XXX move these */ -#define MAX_WIDTH 2048 -#define MAX_HEIGHT 2048 - - -struct llvmpipe_tile_cache -{ - struct pipe_screen *screen; - struct pipe_surface *surface; /**< the surface we're caching */ - struct pipe_transfer *transfer; - void *transfer_map; - - struct llvmpipe_cached_tile entries[MAX_WIDTH/TILE_SIZE][MAX_HEIGHT/TILE_SIZE]; - - uint8_t clear_color[4]; /**< for color bufs */ - uint clear_val; /**< for z+stencil, or packed color clear value */ - - struct llvmpipe_cached_tile *last_tile; /**< most recently retrieved tile */ -}; +struct llvmpipe_tile_cache; /* opaque */ extern struct llvmpipe_tile_cache * diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.c b/src/gallium/drivers/llvmpipe/lp_tile_soa.c deleted file mode 100644 index 4e4ccb31cc..0000000000 --- a/src/gallium/drivers/llvmpipe/lp_tile_soa.c +++ /dev/null @@ -1,931 +0,0 @@ -/************************************************************************** - * - * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. - * All Rights Reserved. - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the - * "Software"), to deal in the Software without restriction, including - * without limitation the rights to use, copy, modify, merge, publish, - * distribute, sub license, and/or sell copies of the Software, and to - * permit persons to whom the Software is furnished to do so, subject to - * the following conditions: - * - * The above copyright notice and this permission notice (including the - * next paragraph) shall be included in all copies or substantial portions - * of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS - * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF - * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. - * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR - * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, - * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE - * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. - * - **************************************************************************/ - -/** - * RGBA/float tile get/put functions. - * Usable both by drivers and state trackers. - */ - - -#include "pipe/p_defines.h" -#include "pipe/p_inlines.h" - -#include "util/u_math.h" -#include "util/u_memory.h" -#include "util/u_rect.h" -#include "util/u_tile.h" -#include "lp_tile_cache.h" -#include "lp_tile_soa.h" - - -const unsigned char -tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = { - { 0, 1, 4, 5, 8, 9, 12, 13}, - { 2, 3, 6, 7, 10, 11, 14, 15} -}; - - - -/*** PIPE_FORMAT_A8R8G8B8_UNORM ***/ - -static void -a8r8g8b8_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const unsigned pixel = *src++; - TILE_PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff; - TILE_PIXEL(p, j, i, 1) = (pixel >> 8) & 0xff; - TILE_PIXEL(p, j, i, 2) = (pixel >> 0) & 0xff; - TILE_PIXEL(p, j, i, 3) = (pixel >> 24) & 0xff; - } - } -} - - -static void -a8r8g8b8_put_tile_rgba(unsigned *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, g, b, a; - r = TILE_PIXEL(p, j, i, 0); - g = TILE_PIXEL(p, j, i, 1); - b = TILE_PIXEL(p, j, i, 2); - a = TILE_PIXEL(p, j, i, 3); - *dst++ = (a << 24) | (r << 16) | (g << 8) | b; - } - } -} - - -/*** PIPE_FORMAT_A8R8G8B8_UNORM ***/ - -static void -x8r8g8b8_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const unsigned pixel = *src++; - TILE_PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff; - TILE_PIXEL(p, j, i, 1) = (pixel >> 8) & 0xff; - TILE_PIXEL(p, j, i, 2) = (pixel >> 0) & 0xff; - TILE_PIXEL(p, j, i, 3) = 0xff; - } - } -} - - -static void -x8r8g8b8_put_tile_rgba(unsigned *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, g, b; - r = TILE_PIXEL(p, j, i, 0); - g = TILE_PIXEL(p, j, i, 1); - b = TILE_PIXEL(p, j, i, 2); - *dst++ = (0xff << 24) | (r << 16) | (g << 8) | b; - } - } -} - - -/*** PIPE_FORMAT_B8G8R8A8_UNORM ***/ - -static void -b8g8r8a8_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const unsigned pixel = *src++; - TILE_PIXEL(p, j, i, 0) = (pixel >> 8) & 0xff; - TILE_PIXEL(p, j, i, 1) = (pixel >> 16) & 0xff; - TILE_PIXEL(p, j, i, 2) = (pixel >> 24) & 0xff; - TILE_PIXEL(p, j, i, 3) = (pixel >> 0) & 0xff; - } - } -} - - -static void -b8g8r8a8_put_tile_rgba(unsigned *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, g, b, a; - r = TILE_PIXEL(p, j, i, 0); - g = TILE_PIXEL(p, j, i, 1); - b = TILE_PIXEL(p, j, i, 2); - a = TILE_PIXEL(p, j, i, 3); - *dst++ = (b << 24) | (g << 16) | (r << 8) | a; - } - } -} - - -/*** PIPE_FORMAT_A1R5G5B5_UNORM ***/ - -static void -a1r5g5b5_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const ushort pixel = *src++; - TILE_PIXEL(p, j, i, 0) = ((pixel >> 10) & 0x1f) * 255 / 31; - TILE_PIXEL(p, j, i, 1) = ((pixel >> 5) & 0x1f) * 255 / 31; - TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0x1f) * 255 / 31; - TILE_PIXEL(p, j, i, 3) = ((pixel >> 15) ) * 255; - } - } -} - - -static void -a1r5g5b5_put_tile_rgba(ushort *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, g, b, a; - r = TILE_PIXEL(p, j, i, 0); - g = TILE_PIXEL(p, j, i, 1); - b = TILE_PIXEL(p, j, i, 2); - a = TILE_PIXEL(p, j, i, 3); - r = r >> 3; /* 5 bits */ - g = g >> 3; /* 5 bits */ - b = b >> 3; /* 5 bits */ - a = a >> 7; /* 1 bit */ - *dst++ = (a << 15) | (r << 10) | (g << 5) | b; - } - } -} - - -/*** PIPE_FORMAT_A4R4G4B4_UNORM ***/ - -static void -a4r4g4b4_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const ushort pixel = *src++; - TILE_PIXEL(p, j, i, 0) = ((pixel >> 8) & 0xf) * 255 / 15; - TILE_PIXEL(p, j, i, 1) = ((pixel >> 4) & 0xf) * 255 / 15; - TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0xf) * 255 / 15; - TILE_PIXEL(p, j, i, 3) = ((pixel >> 12) ) * 255 / 15; - } - } -} - - -static void -a4r4g4b4_put_tile_rgba(ushort *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, g, b, a; - r = TILE_PIXEL(p, j, i, 0); - g = TILE_PIXEL(p, j, i, 1); - b = TILE_PIXEL(p, j, i, 2); - a = TILE_PIXEL(p, j, i, 3); - r >>= 4; - g >>= 4; - b >>= 4; - a >>= 4; - *dst++ = (a << 12) | (r << 16) | (g << 4) | b; - } - } -} - - -/*** PIPE_FORMAT_R5G6B5_UNORM ***/ - -static void -r5g6b5_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const ushort pixel = *src++; - TILE_PIXEL(p, j, i, 0) = ((pixel >> 11) & 0x1f) * 255 / 31; - TILE_PIXEL(p, j, i, 1) = ((pixel >> 5) & 0x3f) * 255 / 63; - TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0x1f) * 255 / 31; - TILE_PIXEL(p, j, i, 3) = 255; - } - } -} - - -static void -r5g6b5_put_tile_rgba(ushort *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - uint r = (uint) TILE_PIXEL(p, j, i, 0) * 31 / 255; - uint g = (uint) TILE_PIXEL(p, j, i, 1) * 63 / 255; - uint b = (uint) TILE_PIXEL(p, j, i, 2) * 31 / 255; - *dst++ = (r << 11) | (g << 5) | (b); - } - } -} - - - -/*** PIPE_FORMAT_Z16_UNORM ***/ - -/** - * Return each Z value as four floats in [0,1]. - */ -static void -z16_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - const float scale = 1.0f / 65535.0f; - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = *src++ * scale; - } - } -} - - - - -/*** PIPE_FORMAT_L8_UNORM ***/ - -static void -l8_get_tile_rgba(const ubyte *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, src++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = *src; - TILE_PIXEL(p, j, i, 3) = 255; - } - } -} - - -static void -l8_put_tile_rgba(ubyte *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r; - r = TILE_PIXEL(p, j, i, 0); - *dst++ = (ubyte) r; - } - } -} - - - -/*** PIPE_FORMAT_A8_UNORM ***/ - -static void -a8_get_tile_rgba(const ubyte *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, src++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = 0; - TILE_PIXEL(p, j, i, 3) = *src; - } - } -} - - -static void -a8_put_tile_rgba(ubyte *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned a; - a = TILE_PIXEL(p, j, i, 3); - *dst++ = (ubyte) a; - } - } -} - - - -/*** PIPE_FORMAT_R16_SNORM ***/ - -static void -r16_get_tile_rgba(const short *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, src++) { - TILE_PIXEL(p, j, i, 0) = MAX2(src[0] >> 7, 0); - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = 0; - TILE_PIXEL(p, j, i, 3) = 255; - } - } -} - - -static void -r16_put_tile_rgba(short *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, dst++) { - dst[0] = TILE_PIXEL(p, j, i, 0) << 7; - } - } -} - - -/*** PIPE_FORMAT_R16G16B16A16_SNORM ***/ - -static void -r16g16b16a16_get_tile_rgba(const short *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, src += 4) { - TILE_PIXEL(p, j, i, 0) = src[0] >> 8; - TILE_PIXEL(p, j, i, 1) = src[1] >> 8; - TILE_PIXEL(p, j, i, 2) = src[2] >> 8; - TILE_PIXEL(p, j, i, 3) = src[3] >> 8; - } - } -} - - -static void -r16g16b16a16_put_tile_rgba(short *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, dst += 4) { - dst[0] = TILE_PIXEL(p, j, i, 0) << 8; - dst[1] = TILE_PIXEL(p, j, i, 1) << 8; - dst[2] = TILE_PIXEL(p, j, i, 2) << 8; - dst[3] = TILE_PIXEL(p, j, i, 3) << 8; - } - } -} - - - -/*** PIPE_FORMAT_I8_UNORM ***/ - -static void -i8_get_tile_rgba(const ubyte *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++, src++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = *src; - } - } -} - - -static void -i8_put_tile_rgba(ubyte *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r; - r = TILE_PIXEL(p, j, i, 0); - *dst++ = (ubyte) r; - } - } -} - - -/*** PIPE_FORMAT_A8L8_UNORM ***/ - -static void -a8l8_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - ushort ra = *src++; - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = ra & 0xff; - TILE_PIXEL(p, j, i, 3) = ra >> 8; - } - } -} - - -static void -a8l8_put_tile_rgba(ushort *dst, - unsigned w, unsigned h, - const uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - unsigned r, a; - r = TILE_PIXEL(p, j, i, 0); - a = TILE_PIXEL(p, j, i, 3); - *dst++ = (a << 8) | r; - } - } -} - - - - -/*** PIPE_FORMAT_Z32_UNORM ***/ - -/** - * Return each Z value as four floats in [0,1]. - */ -static void -z32_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - const double scale = 1.0 / (double) 0xffffffff; - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = (float) (*src++ * scale); - } - } -} - - -/*** PIPE_FORMAT_S8Z24_UNORM ***/ - -/** - * Return Z component as four float in [0,1]. Stencil part ignored. - */ -static void -s8z24_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - const double scale = 1.0 / ((1 << 24) - 1); - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = (float) (scale * (*src++ & 0xffffff)); - } - } -} - - -/*** PIPE_FORMAT_Z24S8_UNORM ***/ - -/** - * Return Z component as four float in [0,1]. Stencil part ignored. - */ -static void -z24s8_get_tile_rgba(const unsigned *src, - unsigned w, unsigned h, - uint8_t *p) -{ - const double scale = 1.0 / ((1 << 24) - 1); - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = (float) (scale * (*src++ >> 8)); - } - } -} - - -/*** PIPE_FORMAT_Z32_FLOAT ***/ - -/** - * Return each Z value as four floats in [0,1]. - */ -static void -z32f_get_tile_rgba(const float *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = *src++; - } - } -} - - -/*** PIPE_FORMAT_YCBCR / PIPE_FORMAT_YCBCR_REV ***/ - -/** - * Convert YCbCr (or YCrCb) to RGBA. - */ -static void -ycbcr_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p, - boolean rev) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - /* do two texels at a time */ - for (j = 0; j < (w & ~1); j += 2, src += 2) { - const ushort t0 = src[0]; - const ushort t1 = src[1]; - const ubyte y0 = (t0 >> 8) & 0xff; /* luminance */ - const ubyte y1 = (t1 >> 8) & 0xff; /* luminance */ - ubyte cb, cr; - float r, g, b; - - if (rev) { - cb = t1 & 0xff; /* chroma U */ - cr = t0 & 0xff; /* chroma V */ - } - else { - cb = t0 & 0xff; /* chroma U */ - cr = t1 & 0xff; /* chroma V */ - } - - /* even pixel: y0,cr,cb */ - r = 1.164f * (y0-16) + 1.596f * (cr-128); - g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128); - b = 1.164f * (y0-16) + 2.018f * (cb-128); - TILE_PIXEL(p, j, i, 0) = r; - TILE_PIXEL(p, j, i, 1) = g; - TILE_PIXEL(p, j, i, 2) = b; - TILE_PIXEL(p, j, i, 3) = 255; - - /* odd pixel: use y1,cr,cb */ - r = 1.164f * (y1-16) + 1.596f * (cr-128); - g = 1.164f * (y1-16) - 0.813f * (cr-128) - 0.391f * (cb-128); - b = 1.164f * (y1-16) + 2.018f * (cb-128); - TILE_PIXEL(p, j + 1, i, 0) = r; - TILE_PIXEL(p, j + 1, i, 1) = g; - TILE_PIXEL(p, j + 1, i, 2) = b; - TILE_PIXEL(p, j + 1, i, 3) = 255; - } - /* do the last texel */ - if (w & 1) { - const ushort t0 = src[0]; - const ushort t1 = src[1]; - const ubyte y0 = (t0 >> 8) & 0xff; /* luminance */ - ubyte cb, cr; - float r, g, b; - - if (rev) { - cb = t1 & 0xff; /* chroma U */ - cr = t0 & 0xff; /* chroma V */ - } - else { - cb = t0 & 0xff; /* chroma U */ - cr = t1 & 0xff; /* chroma V */ - } - - /* even pixel: y0,cr,cb */ - r = 1.164f * (y0-16) + 1.596f * (cr-128); - g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128); - b = 1.164f * (y0-16) + 2.018f * (cb-128); - TILE_PIXEL(p, j, i, 0) = r; - TILE_PIXEL(p, j, i, 1) = g; - TILE_PIXEL(p, j, i, 2) = b; - TILE_PIXEL(p, j, i, 3) = 255; - } - } -} - - -static void -fake_get_tile_rgba(const ushort *src, - unsigned w, unsigned h, - uint8_t *p) -{ - unsigned i, j; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - TILE_PIXEL(p, j, i, 0) = - TILE_PIXEL(p, j, i, 1) = - TILE_PIXEL(p, j, i, 2) = - TILE_PIXEL(p, j, i, 3) = (i ^ j) & 1 ? 255 : 0; - } - } -} - - -static void -lp_tile_raw_to_rgba_soa(enum pipe_format format, - void *src, - uint w, uint h, - uint8_t *p) -{ - switch (format) { - case PIPE_FORMAT_A8R8G8B8_UNORM: - a8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_X8R8G8B8_UNORM: - x8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_B8G8R8A8_UNORM: - b8g8r8a8_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_A1R5G5B5_UNORM: - a1r5g5b5_get_tile_rgba((ushort *) src, w, h, p); - break; - case PIPE_FORMAT_A4R4G4B4_UNORM: - a4r4g4b4_get_tile_rgba((ushort *) src, w, h, p); - break; - case PIPE_FORMAT_R5G6B5_UNORM: - r5g6b5_get_tile_rgba((ushort *) src, w, h, p); - break; - case PIPE_FORMAT_L8_UNORM: - l8_get_tile_rgba((ubyte *) src, w, h, p); - break; - case PIPE_FORMAT_A8_UNORM: - a8_get_tile_rgba((ubyte *) src, w, h, p); - break; - case PIPE_FORMAT_I8_UNORM: - i8_get_tile_rgba((ubyte *) src, w, h, p); - break; - case PIPE_FORMAT_A8L8_UNORM: - a8l8_get_tile_rgba((ushort *) src, w, h, p); - break; - case PIPE_FORMAT_R16_SNORM: - r16_get_tile_rgba((short *) src, w, h, p); - break; - case PIPE_FORMAT_R16G16B16A16_SNORM: - r16g16b16a16_get_tile_rgba((short *) src, w, h, p); - break; - case PIPE_FORMAT_Z16_UNORM: - z16_get_tile_rgba((ushort *) src, w, h, p); - break; - case PIPE_FORMAT_Z32_UNORM: - z32_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_S8Z24_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: - s8z24_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_Z24S8_UNORM: - case PIPE_FORMAT_Z24X8_UNORM: - z24s8_get_tile_rgba((unsigned *) src, w, h, p); - break; - case PIPE_FORMAT_Z32_FLOAT: - z32f_get_tile_rgba((float *) src, w, h, p); - break; - case PIPE_FORMAT_YCBCR: - ycbcr_get_tile_rgba((ushort *) src, w, h, p, FALSE); - break; - case PIPE_FORMAT_YCBCR_REV: - ycbcr_get_tile_rgba((ushort *) src, w, h, p, TRUE); - break; - default: - debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(format)); - fake_get_tile_rgba(src, w, h, p); - } -} - - -void -lp_get_tile_rgba_soa(struct pipe_transfer *pt, - uint x, uint y, - uint8_t *p) -{ - uint w = TILE_SIZE, h = TILE_SIZE; - void *packed; - - if (pipe_clip_tile(x, y, &w, &h, pt)) - return; - - packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size); - - if (!packed) - return; - - if(pt->format == PIPE_FORMAT_YCBCR || pt->format == PIPE_FORMAT_YCBCR_REV) - assert((x & 1) == 0); - - pipe_get_tile_raw(pt, x, y, w, h, packed, 0); - - lp_tile_raw_to_rgba_soa(pt->format, packed, w, h, p); - - FREE(packed); -} - - -void -lp_put_tile_rgba_soa(struct pipe_transfer *pt, - uint x, uint y, - const uint8_t *p) -{ - uint w = TILE_SIZE, h = TILE_SIZE; - void *packed; - - if (pipe_clip_tile(x, y, &w, &h, pt)) - return; - - packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size); - - if (!packed) - return; - - switch (pt->format) { - case PIPE_FORMAT_A8R8G8B8_UNORM: - a8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p); - break; - case PIPE_FORMAT_X8R8G8B8_UNORM: - x8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p); - break; - case PIPE_FORMAT_B8G8R8A8_UNORM: - b8g8r8a8_put_tile_rgba((unsigned *) packed, w, h, p); - break; - case PIPE_FORMAT_A1R5G5B5_UNORM: - a1r5g5b5_put_tile_rgba((ushort *) packed, w, h, p); - break; - case PIPE_FORMAT_R5G6B5_UNORM: - r5g6b5_put_tile_rgba((ushort *) packed, w, h, p); - break; - case PIPE_FORMAT_R8G8B8A8_UNORM: - assert(0); - break; - case PIPE_FORMAT_A4R4G4B4_UNORM: - a4r4g4b4_put_tile_rgba((ushort *) packed, w, h, p); - break; - case PIPE_FORMAT_L8_UNORM: - l8_put_tile_rgba((ubyte *) packed, w, h, p); - break; - case PIPE_FORMAT_A8_UNORM: - a8_put_tile_rgba((ubyte *) packed, w, h, p); - break; - case PIPE_FORMAT_I8_UNORM: - i8_put_tile_rgba((ubyte *) packed, w, h, p); - break; - case PIPE_FORMAT_A8L8_UNORM: - a8l8_put_tile_rgba((ushort *) packed, w, h, p); - break; - case PIPE_FORMAT_R16_SNORM: - r16_put_tile_rgba((short *) packed, w, h, p); - break; - case PIPE_FORMAT_R16G16B16A16_SNORM: - r16g16b16a16_put_tile_rgba((short *) packed, w, h, p); - break; - case PIPE_FORMAT_Z16_UNORM: - /*z16_put_tile_rgba((ushort *) packed, w, h, p);*/ - break; - case PIPE_FORMAT_Z32_UNORM: - /*z32_put_tile_rgba((unsigned *) packed, w, h, p);*/ - break; - case PIPE_FORMAT_S8Z24_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: - /*s8z24_put_tile_rgba((unsigned *) packed, w, h, p);*/ - break; - case PIPE_FORMAT_Z24S8_UNORM: - case PIPE_FORMAT_Z24X8_UNORM: - /*z24s8_put_tile_rgba((unsigned *) packed, w, h, p);*/ - break; - default: - debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(pt->format)); - } - - pipe_put_tile_raw(pt, x, y, w, h, packed, 0); - - FREE(packed); -} - - diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.h b/src/gallium/drivers/llvmpipe/lp_tile_soa.h index 3d8c703b73..040b01865d 100644 --- a/src/gallium/drivers/llvmpipe/lp_tile_soa.h +++ b/src/gallium/drivers/llvmpipe/lp_tile_soa.h @@ -64,14 +64,18 @@ tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH]; void -lp_get_tile_rgba_soa(struct pipe_transfer *pt, - uint x, uint y, - uint8_t *p); +lp_tile_read_4ub(enum pipe_format format, + uint8_t *dst, + const void *src, unsigned src_stride, + unsigned x, unsigned y, unsigned w, unsigned h); + void -lp_put_tile_rgba_soa(struct pipe_transfer *pt, - uint x, uint y, - const uint8_t *p); +lp_tile_write_4ub(enum pipe_format format, + const uint8_t *src, + void *dst, unsigned dst_stride, + unsigned x, unsigned y, unsigned w, unsigned h); + #ifdef __cplusplus diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.py b/src/gallium/drivers/llvmpipe/lp_tile_soa.py new file mode 100644 index 0000000000..004c5c979e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tile_soa.py @@ -0,0 +1,278 @@ +#!/usr/bin/env python + +''' +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Pixel format accessor functions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ +''' + + +import sys +import os.path + +sys.path.insert(0, os.path.join(os.path.dirname(sys.argv[0]), '../../auxiliary/util')) + +from u_format_access import * + + +def generate_format_read(format, dst_type, dst_native_type, dst_suffix): + '''Generate the function to read pixels from a particular format''' + + name = short_name(format) + + src_native_type = native_type(format) + + print 'static void' + print 'lp_tile_%s_read_%s(%s *dst, const uint8_t *src, unsigned src_stride, unsigned x0, unsigned y0, unsigned w, unsigned h)' % (name, dst_suffix, dst_native_type) + print '{' + print ' unsigned x, y;' + print ' const uint8_t *src_row = src + y0*src_stride;' + print ' for (y = 0; y < h; ++y) {' + print ' const %s *src_pixel = (const %s *)(src_row + x0*%u);' % (src_native_type, src_native_type, format.stride()) + print ' for (x = 0; x < w; ++x) {' + + names = ['']*4 + if format.colorspace == 'rgb': + for i in range(4): + swizzle = format.out_swizzle[i] + if swizzle < 4: + names[swizzle] += 'rgba'[i] + elif format.colorspace == 'zs': + swizzle = format.out_swizzle[0] + if swizzle < 4: + names[swizzle] = 'z' + else: + assert False + else: + assert False + + if format.layout == ARITH: + print ' %s pixel = *src_pixel++;' % src_native_type + shift = 0; + for i in range(4): + src_type = format.in_types[i] + width = src_type.size + if names[i]: + value = 'pixel' + mask = (1 << width) - 1 + if shift: + value = '(%s >> %u)' % (value, shift) + if shift + width < format.block_size(): + value = '(%s & 0x%x)' % (value, mask) + value = conversion_expr(src_type, dst_type, dst_native_type, value) + print ' %s %s = %s;' % (dst_native_type, names[i], value) + shift += width + elif format.layout == ARRAY: + for i in range(4): + src_type = format.in_types[i] + if names[i]: + value = '(*src_pixel++)' + value = conversion_expr(src_type, dst_type, dst_native_type, value) + print ' %s %s = %s;' % (dst_native_type, names[i], value) + else: + assert False + + for i in range(4): + if format.colorspace == 'rgb': + swizzle = format.out_swizzle[i] + if swizzle < 4: + value = names[swizzle] + elif swizzle == SWIZZLE_0: + value = '0' + elif swizzle == SWIZZLE_1: + value = '1' + else: + assert False + elif format.colorspace == 'zs': + if i < 3: + value = 'z' + else: + value = '1' + else: + assert False + print ' TILE_PIXEL(dst, x, y, %u) = %s; /* %s */' % (i, value, 'rgba'[i]) + + print ' }' + print ' src_row += src_stride;' + print ' }' + print '}' + print + + +def generate_format_write(format, src_type, src_native_type, src_suffix): + '''Generate the function to write pixels to a particular format''' + + name = short_name(format) + + dst_native_type = native_type(format) + + print 'static void' + print 'lp_tile_%s_write_%s(const %s *src, uint8_t *dst, unsigned dst_stride, unsigned x0, unsigned y0, unsigned w, unsigned h)' % (name, src_suffix, src_native_type) + print '{' + print ' unsigned x, y;' + print ' uint8_t *dst_row = dst + y0*dst_stride;' + print ' for (y = 0; y < h; ++y) {' + print ' %s *dst_pixel = (%s *)(dst_row + x0*%u);' % (dst_native_type, dst_native_type, format.stride()) + print ' for (x = 0; x < w; ++x) {' + + inv_swizzle = [None]*4 + if format.colorspace == 'rgb': + for i in range(4): + swizzle = format.out_swizzle[i] + if swizzle < 4: + inv_swizzle[swizzle] = i + elif format.colorspace == 'zs': + swizzle = format.out_swizzle[0] + if swizzle < 4: + inv_swizzle[swizzle] = 0 + else: + assert False + + if format.layout == ARITH: + print ' %s pixel = 0;' % dst_native_type + shift = 0; + for i in range(4): + dst_type = format.in_types[i] + width = dst_type.size + if inv_swizzle[i] is not None: + value = 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle[i] + value = conversion_expr(src_type, dst_type, dst_native_type, value) + if shift: + value = '(%s << %u)' % (value, shift) + print ' pixel |= %s;' % value + shift += width + print ' *dst_pixel++ = pixel;' + elif format.layout == ARRAY: + for i in range(4): + dst_type = format.in_types[i] + if inv_swizzle[i] is not None: + value = 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle[i] + value = conversion_expr(src_type, dst_type, dst_native_type, value) + print ' *dst_pixel++ = %s;' % value + else: + assert False + + print ' }' + print ' dst_row += dst_stride;' + print ' }' + print '}' + print + + +def generate_read(formats, dst_type, dst_native_type, dst_suffix): + '''Generate the dispatch function to read pixels from any format''' + + for format in formats: + if is_format_supported(format): + generate_format_read(format, dst_type, dst_native_type, dst_suffix) + + print 'void' + print 'lp_tile_read_%s(enum pipe_format format, %s *dst, const void *src, unsigned src_stride, unsigned x, unsigned y, unsigned w, unsigned h)' % (dst_suffix, dst_native_type) + print '{' + print ' void (*func)(%s *dst, const uint8_t *src, unsigned src_stride, unsigned x0, unsigned y0, unsigned w, unsigned h);' % dst_native_type + print ' switch(format) {' + for format in formats: + if is_format_supported(format): + print ' case %s:' % format.name + print ' func = &lp_tile_%s_read_%s;' % (short_name(format), dst_suffix) + print ' break;' + print ' default:' + print ' debug_printf("unsupported format\\n");' + print ' return;' + print ' }' + print ' func(dst, (const uint8_t *)src, src_stride, x, y, w, h);' + print '}' + print + + +def generate_write(formats, src_type, src_native_type, src_suffix): + '''Generate the dispatch function to write pixels to any format''' + + for format in formats: + if is_format_supported(format): + generate_format_write(format, src_type, src_native_type, src_suffix) + + print 'void' + print 'lp_tile_write_%s(enum pipe_format format, const %s *src, void *dst, unsigned dst_stride, unsigned x, unsigned y, unsigned w, unsigned h)' % (src_suffix, src_native_type) + + print '{' + print ' void (*func)(const %s *src, uint8_t *dst, unsigned dst_stride, unsigned x0, unsigned y0, unsigned w, unsigned h);' % src_native_type + print ' switch(format) {' + for format in formats: + if is_format_supported(format): + print ' case %s:' % format.name + print ' func = &lp_tile_%s_write_%s;' % (short_name(format), src_suffix) + print ' break;' + print ' default:' + print ' debug_printf("unsupported format\\n");' + print ' return;' + print ' }' + print ' func(src, (uint8_t *)dst, dst_stride, x, y, w, h);' + print '}' + print + + +def main(): + formats = [] + for arg in sys.argv[1:]: + formats.extend(parse(arg)) + + print '/* This file is autogenerated by lp_tile_soa.py from u_format.csv. Do not edit directly. */' + print + # This will print the copyright message on the top of this file + print __doc__.strip() + print + print '#include "pipe/p_compiler.h"' + print '#include "util/u_format.h"' + print '#include "util/u_math.h"' + print '#include "lp_tile_soa.h"' + print + print 'const unsigned char' + print 'tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = {' + print ' { 0, 1, 4, 5, 8, 9, 12, 13},' + print ' { 2, 3, 6, 7, 10, 11, 14, 15}' + print '};' + print + + generate_clamp() + + type = Type(UNSIGNED, True, 8) + native_type = 'uint8_t' + suffix = '4ub' + + generate_read(formats, type, native_type, suffix) + generate_write(formats, type, native_type, suffix) + + +if __name__ == '__main__': + main() |