diff options
author | José Fonseca <jfonseca@vmware.com> | 2010-01-08 15:42:57 +0000 |
---|---|---|
committer | José Fonseca <jfonseca@vmware.com> | 2010-01-08 15:42:57 +0000 |
commit | 080c40ab32b2abd6d8381b4a0cc143d36a1652b2 (patch) | |
tree | e173767ebc5a82d81b9fc086449d915e29348976 /src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c | |
parent | 9cdf6f025b2ed55cfb13dd09f870f01d0c7947d3 (diff) | |
parent | a1de400e8de06a80ab140bb0fa950e990607572d (diff) |
Merge remote branch 'origin/master' into lp-binning
Conflicts:
src/gallium/auxiliary/util/u_surface.c
src/gallium/drivers/llvmpipe/Makefile
src/gallium/drivers/llvmpipe/SConscript
src/gallium/drivers/llvmpipe/lp_bld_arit.c
src/gallium/drivers/llvmpipe/lp_bld_flow.c
src/gallium/drivers/llvmpipe/lp_bld_interp.c
src/gallium/drivers/llvmpipe/lp_clear.c
src/gallium/drivers/llvmpipe/lp_context.c
src/gallium/drivers/llvmpipe/lp_context.h
src/gallium/drivers/llvmpipe/lp_draw_arrays.c
src/gallium/drivers/llvmpipe/lp_jit.c
src/gallium/drivers/llvmpipe/lp_jit.h
src/gallium/drivers/llvmpipe/lp_prim_vbuf.c
src/gallium/drivers/llvmpipe/lp_setup.c
src/gallium/drivers/llvmpipe/lp_setup_point.c
src/gallium/drivers/llvmpipe/lp_state.h
src/gallium/drivers/llvmpipe/lp_state_blend.c
src/gallium/drivers/llvmpipe/lp_state_derived.c
src/gallium/drivers/llvmpipe/lp_state_fs.c
src/gallium/drivers/llvmpipe/lp_state_sampler.c
src/gallium/drivers/llvmpipe/lp_state_surface.c
src/gallium/drivers/llvmpipe/lp_tex_cache.c
src/gallium/drivers/llvmpipe/lp_tex_cache.h
src/gallium/drivers/llvmpipe/lp_tex_sample.h
src/gallium/drivers/llvmpipe/lp_tile_cache.c
Diffstat (limited to 'src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c')
-rw-r--r-- | src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c | 374 |
1 files changed, 276 insertions, 98 deletions
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c index 1a47ca32d2..5ee8d556a6 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c @@ -27,7 +27,7 @@ /** * @file - * Texture sampling. + * Texture sampling -- SoA. * * @author Jose Fonseca <jfonseca@vmware.com> */ @@ -35,54 +35,23 @@ #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" -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; -} - - - /** * Keep all information for sampling code generation in a single place. */ @@ -111,66 +80,61 @@ struct lp_build_sample_context static void -lp_build_sample_texel(struct lp_build_sample_context *bld, - LLVMValueRef x, - LLVMValueRef y, - LLVMValueRef y_stride, - LLVMValueRef data_ptr, - LLVMValueRef *texel) +lp_build_sample_texel_soa(struct lp_build_sample_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr, + LLVMValueRef *texel) { - struct lp_build_context *int_coord_bld = &bld->int_coord_bld; - LLVMValueRef x_stride; 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); +} - x_stride = lp_build_const_scalar(bld->int_coord_type, bld->format_desc->block.bits/8); - - if(bld->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, int_coord_bld->one, ""); - y_lo = LLVMBuildAnd(bld->builder, y, int_coord_bld->one, ""); - - x_hi = LLVMBuildLShr(bld->builder, x, int_coord_bld->one, ""); - y_hi = LLVMBuildLShr(bld->builder, y, int_coord_bld->one, ""); - - x_stride_lo = x_stride; - y_stride_lo = lp_build_const_scalar(bld->int_coord_type, 2*bld->format_desc->block.bits/8); - - x_stride_hi = lp_build_const_scalar(bld->int_coord_type, 4*bld->format_desc->block.bits/8); - y_stride_hi = LLVMBuildShl(bld->builder, y_stride, int_coord_bld->one, ""); - - x_offset_lo = lp_build_mul(int_coord_bld, x_lo, x_stride_lo); - y_offset_lo = lp_build_mul(int_coord_bld, y_lo, y_stride_lo); - offset_lo = lp_build_add(int_coord_bld, x_offset_lo, y_offset_lo); - - x_offset_hi = lp_build_mul(int_coord_bld, x_hi, x_stride_hi); - y_offset_hi = lp_build_mul(int_coord_bld, y_hi, y_stride_hi); - offset_hi = lp_build_add(int_coord_bld, x_offset_hi, y_offset_hi); - offset = lp_build_add(int_coord_bld, offset_hi, offset_lo); - } - else { - LLVMValueRef x_offset; - LLVMValueRef y_offset; +static LLVMValueRef +lp_build_sample_packed(struct lp_build_sample_context *bld, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef y_stride, + LLVMValueRef data_ptr) +{ + LLVMValueRef offset; - x_offset = lp_build_mul(int_coord_bld, x, x_stride); - y_offset = lp_build_mul(int_coord_bld, y, y_stride); + offset = lp_build_sample_offset(&bld->int_coord_bld, + bld->format_desc, + x, y, y_stride, + data_ptr); - offset = lp_build_add(int_coord_bld, x_offset, y_offset); - } + assert(bld->format_desc->block.width == 1); + assert(bld->format_desc->block.height == 1); + assert(bld->format_desc->block.bits <= bld->texel_type.width); - lp_build_load_rgba_soa(bld->builder, - bld->format_desc, - bld->texel_type, - data_ptr, - offset, - texel); + return lp_build_gather(bld->builder, + bld->texel_type.length, + bld->format_desc->block.bits, + bld->texel_type.width, + data_ptr, offset); } @@ -208,7 +172,8 @@ lp_build_sample_wrap(struct lp_build_sample_context *bld, 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 %u\n", wrap_mode); + _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; @@ -240,7 +205,7 @@ lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld, 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(bld, x, y, stride, data_ptr, texel); + lp_build_sample_texel_soa(bld, x, y, stride, data_ptr, texel); } @@ -286,10 +251,10 @@ lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld, 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(bld, x0, y0, stride, data_ptr, neighbors[0][0]); - lp_build_sample_texel(bld, x1, y0, stride, data_ptr, neighbors[0][1]); - lp_build_sample_texel(bld, x0, y1, stride, data_ptr, neighbors[1][0]); - lp_build_sample_texel(bld, x1, y1, stride, data_ptr, neighbors[1][1]); + 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) { @@ -304,6 +269,217 @@ lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld, 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) @@ -312,7 +488,7 @@ lp_build_sample_compare(struct lp_build_sample_context *bld, LLVMValueRef res; unsigned chan; - if(!bld->static_state->compare_mode) + if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE) return; /* TODO: Compare before swizzling, to avoid redundant computations */ @@ -401,8 +577,10 @@ lp_build_sample_soa(LLVMBuilderRef builder, 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: - lp_build_sample_2d_linear_soa(&bld, s, t, width, height, stride, data_ptr, texel); + 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); |