diff options
Diffstat (limited to 'src/gallium/auxiliary/tgsi')
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_sanity.c | 2 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_sanity.h | 2 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_scan.c | 1 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_scan.h | 1 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_sse2.c | 90 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_ureg.c | 249 | ||||
-rw-r--r-- | src/gallium/auxiliary/tgsi/tgsi_ureg.h | 35 |
7 files changed, 283 insertions, 97 deletions
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sanity.c b/src/gallium/auxiliary/tgsi/tgsi_sanity.c index 8a13885da9..53e13b30e6 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_sanity.c +++ b/src/gallium/auxiliary/tgsi/tgsi_sanity.c @@ -358,7 +358,7 @@ epilog( boolean tgsi_sanity_check( - struct tgsi_token *tokens ) + const struct tgsi_token *tokens ) { struct sanity_check_ctx ctx; diff --git a/src/gallium/auxiliary/tgsi/tgsi_sanity.h b/src/gallium/auxiliary/tgsi/tgsi_sanity.h index ca45e94c7a..52263ff883 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_sanity.h +++ b/src/gallium/auxiliary/tgsi/tgsi_sanity.h @@ -40,7 +40,7 @@ extern "C" { */ boolean tgsi_sanity_check( - struct tgsi_token *tokens ); + const struct tgsi_token *tokens ); #if defined __cplusplus } diff --git a/src/gallium/auxiliary/tgsi/tgsi_scan.c b/src/gallium/auxiliary/tgsi/tgsi_scan.c index c535788819..0db4481a3d 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_scan.c +++ b/src/gallium/auxiliary/tgsi/tgsi_scan.c @@ -132,6 +132,7 @@ tgsi_scan_shader(const struct tgsi_token *tokens, if (file == TGSI_FILE_INPUT) { info->input_semantic_name[reg] = (ubyte)fulldecl->Semantic.SemanticName; info->input_semantic_index[reg] = (ubyte)fulldecl->Semantic.SemanticIndex; + info->input_interpolate[reg] = (ubyte)fulldecl->Declaration.Interpolate; info->num_inputs++; } else if (file == TGSI_FILE_OUTPUT) { diff --git a/src/gallium/auxiliary/tgsi/tgsi_scan.h b/src/gallium/auxiliary/tgsi/tgsi_scan.h index 2c1a75bc81..8a7ee0c7e4 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_scan.h +++ b/src/gallium/auxiliary/tgsi/tgsi_scan.h @@ -45,6 +45,7 @@ struct tgsi_shader_info ubyte num_outputs; ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS]; /**< TGSI_SEMANTIC_x */ ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS]; + ubyte input_interpolate[PIPE_MAX_SHADER_INPUTS]; ubyte output_semantic_name[PIPE_MAX_SHADER_OUTPUTS]; /**< TGSI_SEMANTIC_x */ ubyte output_semantic_index[PIPE_MAX_SHADER_OUTPUTS]; diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c index 1e719940ec..5f6b83b236 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c +++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c @@ -39,8 +39,9 @@ #include "tgsi/tgsi_info.h" #include "tgsi/tgsi_parse.h" #include "tgsi/tgsi_util.h" -#include "tgsi_exec.h" -#include "tgsi_sse2.h" +#include "tgsi/tgsi_dump.h" +#include "tgsi/tgsi_exec.h" +#include "tgsi/tgsi_sse2.h" #include "rtasm/rtasm_x86sse.h" @@ -1360,6 +1361,32 @@ emit_store( const struct tgsi_full_instruction *inst, unsigned chan_index ) { + switch( inst->Instruction.Saturate ) { + case TGSI_SAT_NONE: + break; + + case TGSI_SAT_ZERO_ONE: + sse_maxps( + func, + make_xmm( xmm ), + get_temp( + TGSI_EXEC_TEMP_00000000_I, + TGSI_EXEC_TEMP_00000000_C ) ); + + sse_minps( + func, + make_xmm( xmm ), + get_temp( + TGSI_EXEC_TEMP_ONE_I, + TGSI_EXEC_TEMP_ONE_C ) ); + break; + + case TGSI_SAT_MINUS_PLUS_ONE: + assert( 0 ); + break; + } + + switch( reg->DstRegister.File ) { case TGSI_FILE_OUTPUT: emit_output( @@ -1388,19 +1415,6 @@ emit_store( default: assert( 0 ); } - - switch( inst->Instruction.Saturate ) { - case TGSI_SAT_NONE: - break; - - case TGSI_SAT_ZERO_ONE: - /* assert( 0 ); */ - break; - - case TGSI_SAT_MINUS_PLUS_ONE: - assert( 0 ); - break; - } } #define STORE( FUNC, INST, XMM, INDEX, CHAN )\ @@ -1747,14 +1761,6 @@ emit_instruction( if (indirect_temp_reference(inst)) return FALSE; - /* we don't handle saturation/clamping yet */ - if (inst->Instruction.Saturate != TGSI_SAT_NONE) - return FALSE; - - /* need to use extra temps to fix SOA dependencies : */ - if (tgsi_check_soa_dependencies(inst)) - return FALSE; - switch (inst->Instruction.Opcode) { case TGSI_OPCODE_ARL: FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { @@ -1768,8 +1774,10 @@ emit_instruction( case TGSI_OPCODE_MOV: case TGSI_OPCODE_SWZ: FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { - FETCH( func, *inst, 0, 0, chan_index ); - STORE( func, *inst, 0, 0, chan_index ); + FETCH( func, *inst, 4 + chan_index, 0, chan_index ); + } + FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { + STORE( func, *inst, 4 + chan_index, 0, chan_index ); } break; @@ -1847,7 +1855,6 @@ emit_instruction( break; case TGSI_OPCODE_RCP: - /* TGSI_OPCODE_RECIP */ FETCH( func, *inst, 0, 0, CHAN_X ); emit_rcp( func, 0, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { @@ -1856,7 +1863,6 @@ emit_instruction( break; case TGSI_OPCODE_RSQ: - /* TGSI_OPCODE_RECIPSQRT */ FETCH( func, *inst, 0, 0, CHAN_X ); emit_abs( func, 0 ); emit_rsqrt( func, 1, 0 ); @@ -1954,7 +1960,6 @@ emit_instruction( break; case TGSI_OPCODE_DP3: - /* TGSI_OPCODE_DOT3 */ FETCH( func, *inst, 0, 0, CHAN_X ); FETCH( func, *inst, 1, 1, CHAN_X ); emit_mul( func, 0, 1 ); @@ -1972,7 +1977,6 @@ emit_instruction( break; case TGSI_OPCODE_DP4: - /* TGSI_OPCODE_DOT4 */ FETCH( func, *inst, 0, 0, CHAN_X ); FETCH( func, *inst, 1, 1, CHAN_X ); emit_mul( func, 0, 1 ); @@ -2043,17 +2047,14 @@ emit_instruction( break; case TGSI_OPCODE_SLT: - /* TGSI_OPCODE_SETLT */ emit_setcc( func, inst, cc_LessThan ); break; case TGSI_OPCODE_SGE: - /* TGSI_OPCODE_SETGE */ emit_setcc( func, inst, cc_NotLessThan ); break; case TGSI_OPCODE_MAD: - /* TGSI_OPCODE_MADD */ FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { FETCH( func, *inst, 0, 0, chan_index ); FETCH( func, *inst, 1, 1, chan_index ); @@ -2283,7 +2284,7 @@ emit_instruction( break; case TGSI_OPCODE_SEQ: - return 0; + emit_setcc( func, inst, cc_Equal ); break; case TGSI_OPCODE_SFL: @@ -2291,7 +2292,7 @@ emit_instruction( break; case TGSI_OPCODE_SGT: - return 0; + emit_setcc( func, inst, cc_NotLessThanEqual ); break; case TGSI_OPCODE_SIN: @@ -2303,11 +2304,11 @@ emit_instruction( break; case TGSI_OPCODE_SLE: - return 0; + emit_setcc( func, inst, cc_LessThanEqual ); break; case TGSI_OPCODE_SNE: - return 0; + emit_setcc( func, inst, cc_NotEqual ); break; case TGSI_OPCODE_STR: @@ -2371,7 +2372,6 @@ emit_instruction( break; case TGSI_OPCODE_SSG: - /* TGSI_OPCODE_SGN */ FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { FETCH( func, *inst, 0, 0, chan_index ); emit_sgn( func, 0, 0 ); @@ -2929,6 +2929,22 @@ tgsi_emit_sse2( parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX ? "vertex shader" : "fragment shader"); } + + if (tgsi_check_soa_dependencies(&parse.FullToken.FullInstruction)) { + uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode; + + /* XXX: we only handle src/dst aliasing in a few opcodes + * currently. Need to use an additional temporay to hold + * the result in the cases where the code is too opaque to + * fix. + */ + if (opcode != TGSI_OPCODE_MOV && + opcode != TGSI_OPCODE_SWZ) { + debug_printf("Warning: src/dst aliasing in instruction" + " is not handled:\n"); + tgsi_dump_instruction(&parse.FullToken.FullInstruction, 1); + } + } break; case TGSI_TOKEN_TYPE_IMMEDIATE: diff --git a/src/gallium/auxiliary/tgsi/tgsi_ureg.c b/src/gallium/auxiliary/tgsi/tgsi_ureg.c index f7096bd8e2..654426a903 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_ureg.c +++ b/src/gallium/auxiliary/tgsi/tgsi_ureg.c @@ -31,6 +31,7 @@ #include "tgsi/tgsi_ureg.h" #include "tgsi/tgsi_info.h" #include "tgsi/tgsi_dump.h" +#include "tgsi/tgsi_sanity.h" #include "util/u_memory.h" #include "util/u_math.h" @@ -70,6 +71,7 @@ struct ureg_tokens { #define UREG_MAX_INPUT PIPE_MAX_ATTRIBS #define UREG_MAX_OUTPUT PIPE_MAX_ATTRIBS +#define UREG_MAX_CONSTANT_RANGE 32 #define UREG_MAX_IMMEDIATE 32 #define UREG_MAX_TEMP 256 #define UREG_MAX_ADDR 2 @@ -86,8 +88,10 @@ struct ureg_program unsigned semantic_name; unsigned semantic_index; unsigned interp; - } input[UREG_MAX_INPUT]; - unsigned nr_inputs; + } fs_input[UREG_MAX_INPUT]; + unsigned nr_fs_inputs; + + unsigned vs_inputs[UREG_MAX_INPUT/32]; struct { unsigned semantic_name; @@ -107,9 +111,13 @@ struct ureg_program unsigned temps_active[UREG_MAX_TEMP / 32]; unsigned nr_temps; - unsigned nr_addrs; + struct { + unsigned first; + unsigned last; + } constant_range[UREG_MAX_CONSTANT_RANGE]; + unsigned nr_constant_ranges; - unsigned nr_constants; + unsigned nr_addrs; unsigned nr_instructions; struct ureg_tokens domain[2]; @@ -119,6 +127,9 @@ static union tgsi_any_token error_tokens[32]; static void tokens_error( struct ureg_tokens *tokens ) { + if (tokens->tokens && tokens->tokens != error_tokens) + FREE(tokens->tokens); + tokens->tokens = error_tokens; tokens->size = Elements(error_tokens); tokens->count = 0; @@ -228,25 +239,25 @@ ureg_src_register( unsigned file, -static struct ureg_src -ureg_DECL_input( struct ureg_program *ureg, - unsigned name, - unsigned index, - unsigned interp_mode ) +struct ureg_src +ureg_DECL_fs_input( struct ureg_program *ureg, + unsigned name, + unsigned index, + unsigned interp_mode ) { unsigned i; - for (i = 0; i < ureg->nr_inputs; i++) { - if (ureg->input[i].semantic_name == name && - ureg->input[i].semantic_index == index) + for (i = 0; i < ureg->nr_fs_inputs; i++) { + if (ureg->fs_input[i].semantic_name == name && + ureg->fs_input[i].semantic_index == index) goto out; } - if (ureg->nr_inputs < UREG_MAX_INPUT) { - ureg->input[i].semantic_name = name; - ureg->input[i].semantic_index = index; - ureg->input[i].interp = interp_mode; - ureg->nr_inputs++; + if (ureg->nr_fs_inputs < UREG_MAX_INPUT) { + ureg->fs_input[i].semantic_name = name; + ureg->fs_input[i].semantic_index = index; + ureg->fs_input[i].interp = interp_mode; + ureg->nr_fs_inputs++; } else { set_bad( ureg ); @@ -257,25 +268,14 @@ out: } - -struct ureg_src -ureg_DECL_fs_input( struct ureg_program *ureg, - unsigned name, - unsigned index, - unsigned interp ) -{ - assert(ureg->processor == TGSI_PROCESSOR_FRAGMENT); - return ureg_DECL_input( ureg, name, index, interp ); -} - - struct ureg_src ureg_DECL_vs_input( struct ureg_program *ureg, - unsigned name, unsigned index ) { assert(ureg->processor == TGSI_PROCESSOR_VERTEX); - return ureg_DECL_input( ureg, name, index, TGSI_INTERPOLATE_CONSTANT ); + + ureg->vs_inputs[index/32] |= 1 << (index % 32); + return ureg_src_register( TGSI_FILE_INPUT, index ); } @@ -313,9 +313,57 @@ out: * value or manage any constant_buffer contents -- that's the * resposibility of the calling code. */ -struct ureg_src ureg_DECL_constant(struct ureg_program *ureg ) +struct ureg_src ureg_DECL_constant(struct ureg_program *ureg, + unsigned index ) { - return ureg_src_register( TGSI_FILE_CONSTANT, ureg->nr_constants++ ); + unsigned minconst = index, maxconst = index; + unsigned i; + + /* Inside existing range? + */ + for (i = 0; i < ureg->nr_constant_ranges; i++) { + if (ureg->constant_range[i].first <= index && + ureg->constant_range[i].last >= index) + goto out; + } + + /* Extend existing range? + */ + for (i = 0; i < ureg->nr_constant_ranges; i++) { + if (ureg->constant_range[i].last == index - 1) { + ureg->constant_range[i].last = index; + goto out; + } + + if (ureg->constant_range[i].first == index + 1) { + ureg->constant_range[i].first = index; + goto out; + } + + minconst = MIN2(minconst, ureg->constant_range[i].first); + maxconst = MAX2(maxconst, ureg->constant_range[i].last); + } + + /* Create new range? + */ + if (ureg->nr_constant_ranges < UREG_MAX_CONSTANT_RANGE) { + i = ureg->nr_constant_ranges++; + ureg->constant_range[i].first = index; + ureg->constant_range[i].last = index; + } + + /* Collapse all ranges down to one: + */ + i = 0; + ureg->constant_range[0].first = minconst; + ureg->constant_range[0].last = maxconst; + ureg->nr_constant_ranges = 1; + +out: + assert(i < ureg->nr_constant_ranges); + assert(ureg->constant_range[i].first <= index); + assert(ureg->constant_range[i].last >= index); + return ureg_src_register( TGSI_FILE_CONSTANT, index ); } @@ -566,6 +614,19 @@ ureg_emit_dst( struct ureg_program *ureg, } +static void validate( unsigned opcode, + unsigned nr_dst, + unsigned nr_src ) +{ +#ifdef DEBUG + const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode ); + assert(info); + if(info) { + assert(nr_dst == info->num_dst); + assert(nr_src == info->num_src); + } +#endif +} unsigned ureg_emit_insn(struct ureg_program *ureg, @@ -576,6 +637,8 @@ ureg_emit_insn(struct ureg_program *ureg, { union tgsi_any_token *out; + validate( opcode, num_dst, num_src ); + out = get_tokens( ureg, DOMAIN_INSN, 1 ); out[0].value = 0; out[0].insn.Type = TGSI_TOKEN_TYPE_INSTRUCTION; @@ -678,17 +741,6 @@ ureg_insn(struct ureg_program *ureg, unsigned insn, i; boolean saturate; -#ifdef DEBUG - { - const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode ); - assert(info); - if(info) { - assert(nr_dst == info->num_dst); - assert(nr_src == info->num_src); - } - } -#endif - saturate = nr_dst ? dst[0].Saturate : FALSE; insn = ureg_emit_insn( ureg, opcode, saturate, nr_dst, nr_src ); @@ -702,6 +754,53 @@ ureg_insn(struct ureg_program *ureg, ureg_fixup_insn_size( ureg, insn ); } +void +ureg_tex_insn(struct ureg_program *ureg, + unsigned opcode, + const struct ureg_dst *dst, + unsigned nr_dst, + unsigned target, + const struct ureg_src *src, + unsigned nr_src ) +{ + unsigned insn, i; + boolean saturate; + + saturate = nr_dst ? dst[0].Saturate : FALSE; + + insn = ureg_emit_insn( ureg, opcode, saturate, nr_dst, nr_src ); + + ureg_emit_texture( ureg, insn, target ); \ + + for (i = 0; i < nr_dst; i++) + ureg_emit_dst( ureg, dst[i] ); + + for (i = 0; i < nr_src; i++) + ureg_emit_src( ureg, src[i] ); + + ureg_fixup_insn_size( ureg, insn ); +} + + +void +ureg_label_insn(struct ureg_program *ureg, + unsigned opcode, + const struct ureg_src *src, + unsigned nr_src, + unsigned *label_token ) +{ + unsigned insn, i; + + insn = ureg_emit_insn( ureg, opcode, FALSE, 0, nr_src ); + + ureg_emit_label( ureg, insn, label_token ); \ + + for (i = 0; i < nr_src; i++) + ureg_emit_src( ureg, src[i] ); + + ureg_fixup_insn_size( ureg, insn ); +} + static void emit_decl( struct ureg_program *ureg, @@ -777,13 +876,22 @@ static void emit_decls( struct ureg_program *ureg ) { unsigned i; - for (i = 0; i < ureg->nr_inputs; i++) { - emit_decl( ureg, - TGSI_FILE_INPUT, - i, - ureg->input[i].semantic_name, - ureg->input[i].semantic_index, - ureg->input[i].interp ); + if (ureg->processor == TGSI_PROCESSOR_VERTEX) { + for (i = 0; i < UREG_MAX_INPUT; i++) { + if (ureg->vs_inputs[i/32] & (1 << (i%32))) { + emit_decl_range( ureg, TGSI_FILE_INPUT, i, 1 ); + } + } + } + else { + for (i = 0; i < ureg->nr_fs_inputs; i++) { + emit_decl( ureg, + TGSI_FILE_INPUT, + i, + ureg->fs_input[i].semantic_name, + ureg->fs_input[i].semantic_index, + ureg->fs_input[i].interp ); + } } for (i = 0; i < ureg->nr_outputs; i++) { @@ -801,10 +909,13 @@ static void emit_decls( struct ureg_program *ureg ) ureg->sampler[i].Index, 1 ); } - if (ureg->nr_constants) { - emit_decl_range( ureg, - TGSI_FILE_CONSTANT, - 0, ureg->nr_constants ); + if (ureg->nr_constant_ranges) { + for (i = 0; i < ureg->nr_constant_ranges; i++) + emit_decl_range( ureg, + TGSI_FILE_CONSTANT, + ureg->constant_range[i].first, + (ureg->constant_range[i].last + 1 - + ureg->constant_range[i].first) ); } if (ureg->nr_temps) { @@ -890,6 +1001,15 @@ const struct tgsi_token *ureg_finalize( struct ureg_program *ureg ) ureg->domain[DOMAIN_DECL].count); tgsi_dump( tokens, 0 ); } + +#if DEBUG + if (tokens && !tgsi_sanity_check(tokens)) { + debug_printf("tgsi_ureg.c, sanity check failed on generated tokens:\n"); + tgsi_dump(tokens, 0); + assert(0); + } +#endif + return tokens; } @@ -911,6 +1031,25 @@ void *ureg_create_shader( struct ureg_program *ureg, } +const struct tgsi_token *ureg_get_tokens( struct ureg_program *ureg, + unsigned *nr_tokens ) +{ + const struct tgsi_token *tokens; + + ureg_finalize(ureg); + + tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token; + + if (nr_tokens) + *nr_tokens = ureg->domain[DOMAIN_DECL].size; + + ureg->domain[DOMAIN_DECL].tokens = 0; + ureg->domain[DOMAIN_DECL].size = 0; + ureg->domain[DOMAIN_DECL].order = 0; + ureg->domain[DOMAIN_DECL].count = 0; + + return tokens; +} struct ureg_program *ureg_create( unsigned processor ) diff --git a/src/gallium/auxiliary/tgsi/tgsi_ureg.h b/src/gallium/auxiliary/tgsi/tgsi_ureg.h index acbca59040..f04f443b9e 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_ureg.h +++ b/src/gallium/auxiliary/tgsi/tgsi_ureg.h @@ -82,10 +82,21 @@ ureg_create( unsigned processor ); const struct tgsi_token * ureg_finalize( struct ureg_program * ); +/* Create and return a shader: + */ void * ureg_create_shader( struct ureg_program *, struct pipe_context *pipe ); + +/* Alternately, return the built token stream and hand ownership of + * that memory to the caller: + */ +const struct tgsi_token * +ureg_get_tokens( struct ureg_program *ureg, + unsigned *nr_tokens ); + + void ureg_destroy( struct ureg_program * ); @@ -116,8 +127,7 @@ ureg_DECL_fs_input( struct ureg_program *, struct ureg_src ureg_DECL_vs_input( struct ureg_program *, - unsigned semantic_name, - unsigned semantic_index ); + unsigned index ); struct ureg_dst ureg_DECL_output( struct ureg_program *, @@ -130,7 +140,8 @@ ureg_DECL_immediate( struct ureg_program *, unsigned nr ); struct ureg_src -ureg_DECL_constant( struct ureg_program * ); +ureg_DECL_constant( struct ureg_program *, + unsigned index ); struct ureg_dst ureg_DECL_temporary( struct ureg_program * ); @@ -233,6 +244,24 @@ ureg_insn(struct ureg_program *ureg, unsigned nr_src ); +void +ureg_tex_insn(struct ureg_program *ureg, + unsigned opcode, + const struct ureg_dst *dst, + unsigned nr_dst, + unsigned target, + const struct ureg_src *src, + unsigned nr_src ); + + +void +ureg_label_insn(struct ureg_program *ureg, + unsigned opcode, + const struct ureg_src *src, + unsigned nr_src, + unsigned *label); + + /*********************************************************************** * Internal instruction helpers, don't call these directly: */ |