X-Git-Url: https://git.libre-soc.org/?p=mesa.git;a=blobdiff_plain;f=src%2Fmesa%2Fstate_tracker%2Fst_mesa_to_tgsi.c;h=42f1c2017f8479a2bdc3ab1b80025d44df9afcd9;hp=06cf6d21c6793b814535c697283a8b9c6a041363;hb=652901e95b4ed406293d0e1fabee857c054119b1;hpb=ae4dde6598eacb64bf865037bdaea2041499aa27 diff --git a/src/mesa/state_tracker/st_mesa_to_tgsi.c b/src/mesa/state_tracker/st_mesa_to_tgsi.c index 06cf6d21c67..42f1c2017f8 100644 --- a/src/mesa/state_tracker/st_mesa_to_tgsi.c +++ b/src/mesa/state_tracker/st_mesa_to_tgsi.c @@ -32,18 +32,28 @@ */ #include "pipe/p_compiler.h" +#include "pipe/p_context.h" +#include "pipe/p_screen.h" #include "pipe/p_shader_tokens.h" #include "pipe/p_state.h" -#include "pipe/p_context.h" #include "tgsi/tgsi_ureg.h" #include "st_mesa_to_tgsi.h" #include "st_context.h" -#include "shader/prog_instruction.h" -#include "shader/prog_parameter.h" +#include "program/prog_instruction.h" +#include "program/prog_parameter.h" #include "util/u_debug.h" #include "util/u_math.h" #include "util/u_memory.h" + +#define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \ + (1 << PROGRAM_ENV_PARAM) | \ + (1 << PROGRAM_STATE_VAR) | \ + (1 << PROGRAM_NAMED_PARAM) | \ + (1 << PROGRAM_CONSTANT) | \ + (1 << PROGRAM_UNIFORM)) + + struct label { unsigned branch_target; unsigned token; @@ -62,6 +72,13 @@ struct st_translate { struct ureg_src inputs[PIPE_MAX_SHADER_INPUTS]; struct ureg_dst address[1]; struct ureg_src samplers[PIPE_MAX_SAMPLERS]; + struct ureg_src systemValues[SYSTEM_VALUE_MAX]; + + /* Extra info for handling point size clamping in vertex shader */ + struct ureg_dst pointSizeResult; /**< Actual point size output register */ + struct ureg_src pointSizeConst; /**< Point size range constant register */ + GLint pointSizeOutIndex; /**< Temp point size output register */ + GLboolean prevInstWrotePointSize; const GLuint *inputMapping; const GLuint *outputMapping; @@ -88,6 +105,19 @@ struct st_translate { }; +/** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */ +static unsigned mesa_sysval_to_semantic[SYSTEM_VALUE_MAX] = { + TGSI_SEMANTIC_FACE, + TGSI_SEMANTIC_INSTANCEID +}; + + +/** + * Make note of a branch to a label in the TGSI code. + * After we've emitted all instructions, we'll go over the list + * of labels built here and patch the TGSI code with the actual + * location of each label. + */ static unsigned *get_label( struct st_translate *t, unsigned branch_target ) { @@ -112,6 +142,12 @@ static unsigned *get_label( struct st_translate *t, } +/** + * Called prior to emitting the TGSI code for each Mesa instruction. + * Allocate additional space for instructions if needed. + * Update the insn[] array so the next Mesa instruction points to + * the next TGSI instruction. + */ static void set_insn_start( struct st_translate *t, unsigned start ) { @@ -131,8 +167,8 @@ static void set_insn_start( struct st_translate *t, } -/* - * Map mesa register file to TGSI register file. +/** + * Map a Mesa dst register to a TGSI ureg_dst register. */ static struct ureg_dst dst_register( struct st_translate *t, @@ -150,6 +186,18 @@ dst_register( struct st_translate *t, return t->temps[index]; case PROGRAM_OUTPUT: + if (t->procType == TGSI_PROCESSOR_VERTEX && index == VERT_RESULT_PSIZ) + t->prevInstWrotePointSize = GL_TRUE; + + if (t->procType == TGSI_PROCESSOR_VERTEX) + assert(index < VERT_RESULT_MAX); + else if (t->procType == TGSI_PROCESSOR_FRAGMENT) + assert(index < FRAG_RESULT_MAX); + else + assert(index < GEOM_RESULT_MAX); + + assert(t->outputMapping[index] < Elements(t->outputs)); + return t->outputs[t->outputMapping[index]]; case PROGRAM_ADDRESS: @@ -162,6 +210,9 @@ dst_register( struct st_translate *t, } +/** + * Map a Mesa src register to a TGSI ureg_src register. + */ static struct ureg_src src_register( struct st_translate *t, gl_register_file file, @@ -172,16 +223,17 @@ src_register( struct st_translate *t, return ureg_src_undef(); case PROGRAM_TEMPORARY: - ASSERT(index >= 0); + assert(index >= 0); if (ureg_dst_is_undef(t->temps[index])) t->temps[index] = ureg_DECL_temporary( t->ureg ); + assert(index < Elements(t->temps)); return ureg_src(t->temps[index]); case PROGRAM_NAMED_PARAM: case PROGRAM_ENV_PARAM: case PROGRAM_LOCAL_PARAM: case PROGRAM_UNIFORM: - ASSERT(index >= 0); + assert(index >= 0); return t->constants[index]; case PROGRAM_STATE_VAR: case PROGRAM_CONSTANT: /* ie, immediate */ @@ -191,14 +243,20 @@ src_register( struct st_translate *t, return t->constants[index]; case PROGRAM_INPUT: + assert(t->inputMapping[index] < Elements(t->inputs)); return t->inputs[t->inputMapping[index]]; case PROGRAM_OUTPUT: + assert(t->outputMapping[index] < Elements(t->outputs)); return ureg_src(t->outputs[t->outputMapping[index]]); /* not needed? */ case PROGRAM_ADDRESS: return ureg_src(t->address[index]); + case PROGRAM_SYSTEM_VALUE: + assert(index < Elements(t->systemValues)); + return t->systemValues[index]; + default: debug_assert( 0 ); return ureg_src_undef(); @@ -235,6 +293,9 @@ translate_texture_target( GLuint textarget, } +/** + * Create a TGSI ureg_dst register from a Mesa dest register. + */ static struct ureg_dst translate_dst( struct st_translate *t, const struct prog_dst_register *DstReg, @@ -257,12 +318,24 @@ translate_dst( struct st_translate *t, } +/** + * Create a TGSI ureg_src register from a Mesa src register. + */ static struct ureg_src translate_src( struct st_translate *t, const struct prog_src_register *SrcReg ) { struct ureg_src src = src_register( t, SrcReg->File, SrcReg->Index ); + if (t->procType == TGSI_PROCESSOR_GEOMETRY && SrcReg->HasIndex2) { + src = src_register( t, SrcReg->File, SrcReg->Index2 ); + if (SrcReg->RelAddr2) + src = ureg_src_dimension_indirect( src, ureg_src(t->address[0]), + SrcReg->Index); + else + src = ureg_src_dimension( src, SrcReg->Index); + } + src = ureg_swizzle( src, GET_SWZ( SrcReg->Swizzle, 0 ) & 0x3, GET_SWZ( SrcReg->Swizzle, 1 ) & 0x3, @@ -277,10 +350,15 @@ translate_src( struct st_translate *t, if (SrcReg->RelAddr) { src = ureg_src_indirect( src, ureg_src(t->address[0])); - /* If SrcReg->Index was negative, it was set to zero in - * src_register(). Reassign it now. - */ - src.Index = SrcReg->Index; + if (SrcReg->File != PROGRAM_INPUT && + SrcReg->File != PROGRAM_OUTPUT) { + /* If SrcReg->Index was negative, it was set to zero in + * src_register(). Reassign it now. But don't do this + * for input/output regs since they get remapped while + * const buffers don't. + */ + src.Index = SrcReg->Index; + } } return src; @@ -475,6 +553,10 @@ translate_opcode( unsigned op ) return TGSI_OPCODE_DST; case OPCODE_ELSE: return TGSI_OPCODE_ELSE; + case OPCODE_EMIT_VERTEX: + return TGSI_OPCODE_EMIT; + case OPCODE_END_PRIMITIVE: + return TGSI_OPCODE_ENDPRIM; case OPCODE_ENDIF: return TGSI_OPCODE_ENDIF; case OPCODE_ENDLOOP: @@ -666,6 +748,7 @@ compile_instruction( } } + /** * Emit the TGSI instructions to adjust the WPOS pixel center convention */ @@ -677,18 +760,25 @@ emit_adjusted_wpos( struct st_translate *t, struct ureg_dst wpos_temp = ureg_DECL_temporary(ureg); struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]]; - ureg_ADD(ureg, ureg_writemask(wpos_temp, TGSI_WRITEMASK_X | TGSI_WRITEMASK_Y), - wpos_input, ureg_imm1f(ureg, value)); + /* Note that we bias X and Y and pass Z and W through unchanged. + * The shader might also use gl_FragCoord.w and .z. + */ + ureg_ADD(ureg, wpos_temp, wpos_input, + ureg_imm4f(ureg, value, value, 0.0f, 0.0f)); t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp); } + /** * Emit the TGSI instructions for inverting the WPOS y coordinate. + * This code is unavoidable because it also depends on whether + * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM). */ static void -emit_inverted_wpos( struct st_translate *t, - const struct gl_program *program ) +emit_wpos_inversion( struct st_translate *t, + const struct gl_program *program, + boolean invert) { struct ureg_program *ureg = t->ureg; @@ -696,17 +786,17 @@ emit_inverted_wpos( struct st_translate *t, * Need to replace instances of INPUT[WPOS] with temp T * where T = INPUT[WPOS] by y is inverted. */ - static const gl_state_index winSizeState[STATE_LENGTH] - = { STATE_INTERNAL, STATE_FB_SIZE, 0, 0, 0 }; + static const gl_state_index wposTransformState[STATE_LENGTH] + = { STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM, 0, 0, 0 }; /* XXX: note we are modifying the incoming shader here! Need to * do this before emitting the constant decls below, or this * will be missed: */ - unsigned winHeightConst = _mesa_add_state_reference(program->Parameters, - winSizeState); + unsigned wposTransConst = _mesa_add_state_reference(program->Parameters, + wposTransformState); - struct ureg_src winsize = ureg_DECL_constant( ureg, winHeightConst ); + struct ureg_src wpostrans = ureg_DECL_constant( ureg, wposTransConst ); struct ureg_dst wpos_temp; struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]]; @@ -719,12 +809,23 @@ emit_inverted_wpos( struct st_translate *t, ureg_MOV( ureg, wpos_temp, wpos_input ); } - /* SUB wpos_temp.y, winsize_const, wpos_input - */ - ureg_SUB( ureg, - ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ), - winsize, - wpos_input); + if (invert) { + /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy + */ + ureg_MAD( ureg, + ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ), + wpos_input, + ureg_scalar(wpostrans, 0), + ureg_scalar(wpostrans, 1)); + } else { + /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww + */ + ureg_MAD( ureg, + ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ), + wpos_input, + ureg_scalar(wpostrans, 2), + ureg_scalar(wpostrans, 3)); + } /* Use wpos_temp as position input from here on: */ @@ -732,6 +833,64 @@ emit_inverted_wpos( struct st_translate *t, } +/** + * Emit fragment position/ooordinate code. + */ +static void +emit_wpos(struct st_context *st, + struct st_translate *t, + const struct gl_program *program, + struct ureg_program *ureg) +{ + const struct gl_fragment_program *fp = + (const struct gl_fragment_program *) program; + struct pipe_screen *pscreen = st->pipe->screen; + boolean invert = FALSE; + + if (fp->OriginUpperLeft) { + if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) { + } + else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) { + ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT); + invert = TRUE; + } + else + assert(0); + } + else { + if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) + ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT); + else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) + invert = TRUE; + else + assert(0); + } + + if (fp->PixelCenterInteger) { + if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) + ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER); + else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) + emit_adjusted_wpos(t, program, invert ? 0.5f : -0.5f); + else + assert(0); + } + else { + if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) { + } + else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) { + ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER); + emit_adjusted_wpos(t, program, invert ? -0.5f : 0.5f); + } + else + assert(0); + } + + /* we invert after adjustment so that we avoid the MOV to temporary, + * and reuse the adjustment ADD instead */ + emit_wpos_inversion(t, program, invert); +} + + /** * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back. * TGSI uses +1 for front, -1 for back. @@ -790,7 +949,7 @@ emit_edgeflags( struct st_translate *t, */ enum pipe_error st_translate_mesa_program( - GLcontext *ctx, + struct gl_context *ctx, uint procType, struct ureg_program *ureg, const struct gl_program *program, @@ -809,6 +968,9 @@ st_translate_mesa_program( unsigned i; enum pipe_error ret = PIPE_OK; + assert(numInputs <= Elements(t->inputs)); + assert(numOutputs <= Elements(t->outputs)); + t = &translate; memset(t, 0, sizeof *t); @@ -816,6 +978,8 @@ st_translate_mesa_program( t->inputMapping = inputMapping; t->outputMapping = outputMapping; t->ureg = ureg; + t->pointSizeOutIndex = -1; + t->prevInstWrotePointSize = GL_FALSE; /*_mesa_print_program(program);*/ @@ -823,7 +987,6 @@ st_translate_mesa_program( * Declare input attributes. */ if (procType == TGSI_PROCESSOR_FRAGMENT) { - struct gl_fragment_program* fp = (struct gl_fragment_program*)program; for (i = 0; i < numInputs; i++) { if (program->InputFlags[0] & PROG_PARAM_BIT_CYL_WRAP) { t->inputs[i] = ureg_DECL_fs_input_cyl(ureg, @@ -844,51 +1007,7 @@ st_translate_mesa_program( /* Must do this after setting up t->inputs, and before * emitting constant references, below: */ - struct pipe_screen* pscreen = st_context(ctx)->pipe->screen; - boolean invert = FALSE; - - if (fp->OriginUpperLeft) { - if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) { - } - else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) { - ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT); - invert = TRUE; - } - else - assert(0); - } - else { - if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) - ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT); - else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) - invert = TRUE; - else - assert(0); - } - - if (fp->PixelCenterInteger) { - if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) - ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER); - else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) - emit_adjusted_wpos(t, program, invert ? 0.5f : -0.5f); - else - assert(0); - } - else { - if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) { - } - else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) { - ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER); - emit_adjusted_wpos(t, program, invert ? -0.5f : 0.5f); - } - else - assert(0); - } - - /* we invert after adjustment so that we avoid the MOV to temporary, - * and reuse the adjustment ADD instead */ - if (invert) - emit_inverted_wpos(t, program); + emit_wpos(st_context(ctx), t, program, ureg); } if (program->InputsRead & FRAG_BIT_FACE) { @@ -908,6 +1027,13 @@ st_translate_mesa_program( t->outputs[i] = ureg_writemask( t->outputs[i], TGSI_WRITEMASK_Z ); break; + case TGSI_SEMANTIC_STENCIL: + t->outputs[i] = ureg_DECL_output( ureg, + TGSI_SEMANTIC_STENCIL, /* Stencil */ + outputSemanticIndex[i] ); + t->outputs[i] = ureg_writemask( t->outputs[i], + TGSI_WRITEMASK_Y ); + break; case TGSI_SEMANTIC_COLOR: t->outputs[i] = ureg_DECL_output( ureg, TGSI_SEMANTIC_COLOR, @@ -919,7 +1045,23 @@ st_translate_mesa_program( } } } + else if (procType == TGSI_PROCESSOR_GEOMETRY) { + for (i = 0; i < numInputs; i++) { + t->inputs[i] = ureg_DECL_gs_input(ureg, + i, + inputSemanticName[i], + inputSemanticIndex[i]); + } + + for (i = 0; i < numOutputs; i++) { + t->outputs[i] = ureg_DECL_output( ureg, + outputSemanticName[i], + outputSemanticIndex[i] ); + } + } else { + assert(procType == TGSI_PROCESSOR_VERTEX); + for (i = 0; i < numInputs; i++) { t->inputs[i] = ureg_DECL_vs_input(ureg, i); } @@ -928,6 +1070,25 @@ st_translate_mesa_program( t->outputs[i] = ureg_DECL_output( ureg, outputSemanticName[i], outputSemanticIndex[i] ); + if ((outputSemanticName[i] == TGSI_SEMANTIC_PSIZE) && program->Id) { + /* Writing to the point size result register requires special + * handling to implement clamping. + */ + static const gl_state_index pointSizeClampState[STATE_LENGTH] + = { STATE_INTERNAL, STATE_POINT_SIZE_IMPL_CLAMP, 0, 0, 0 }; + /* XXX: note we are modifying the incoming shader here! Need to + * do this before emitting the constant decls below, or this + * will be missed: + */ + unsigned pointSizeClampConst = + _mesa_add_state_reference(program->Parameters, + pointSizeClampState); + struct ureg_dst psizregtemp = ureg_DECL_temporary( ureg ); + t->pointSizeConst = ureg_DECL_constant( ureg, pointSizeClampConst ); + t->pointSizeResult = t->outputs[i]; + t->pointSizeOutIndex = i; + t->outputs[i] = psizregtemp; + } } if (passthrough_edgeflags) emit_edgeflags( t, program ); @@ -940,19 +1101,42 @@ st_translate_mesa_program( t->address[0] = ureg_DECL_address( ureg ); } + /* Declare misc input registers + */ + { + GLbitfield sysInputs = program->SystemValuesRead; + unsigned numSys = 0; + for (i = 0; sysInputs; i++) { + if (sysInputs & (1 << i)) { + unsigned semName = mesa_sysval_to_semantic[i]; + t->systemValues[i] = ureg_DECL_system_value(ureg, numSys, semName, 0); + numSys++; + sysInputs &= ~(1 << i); + } + } + } + + if (program->IndirectRegisterFiles & (1 << PROGRAM_TEMPORARY)) { + /* If temps are accessed with indirect addressing, declare temporaries + * in sequential order. Else, we declare them on demand elsewhere. + */ + for (i = 0; i < program->NumTemporaries; i++) { + /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */ + t->temps[i] = ureg_DECL_temporary( t->ureg ); + } + } /* Emit constants and immediates. Mesa uses a single index space * for these, so we put all the translated regs in t->constants. */ if (program->Parameters) { - t->constants = CALLOC( program->Parameters->NumParameters, sizeof t->constants[0] ); if (t->constants == NULL) { ret = PIPE_ERROR_OUT_OF_MEMORY; goto out; } - + for (i = 0; i < program->Parameters->NumParameters; i++) { switch (program->Parameters->Parameters[i].Type) { case PROGRAM_ENV_PARAM: @@ -963,13 +1147,14 @@ st_translate_mesa_program( t->constants[i] = ureg_DECL_constant( ureg, i ); break; - /* Emit immediates only when there is no address register - * in use. FIXME: Be smarter and recognize param arrays: + /* Emit immediates only when there's no indirect addressing of + * the const buffer. + * FIXME: Be smarter and recognize param arrays: * indirect addressing is only valid within the referenced * array. */ case PROGRAM_CONSTANT: - if (program->NumAddressRegs > 0) + if (program->IndirectRegisterFiles & PROGRAM_ANY_CONST) t->constants[i] = ureg_DECL_constant( ureg, i ); else t->constants[i] = @@ -995,6 +1180,25 @@ st_translate_mesa_program( for (i = 0; i < program->NumInstructions; i++) { set_insn_start( t, ureg_get_instruction_number( ureg )); compile_instruction( t, &program->Instructions[i] ); + + if (t->prevInstWrotePointSize && program->Id) { + /* The previous instruction wrote to the (fake) vertex point size + * result register. Now we need to clamp that value to the min/max + * point size range, putting the result into the real point size + * register. + * Note that we can't do this easily at the end of program due to + * possible early return. + */ + set_insn_start( t, ureg_get_instruction_number( ureg )); + ureg_MAX( t->ureg, + ureg_writemask(t->outputs[t->pointSizeOutIndex], WRITEMASK_X), + ureg_src(t->outputs[t->pointSizeOutIndex]), + ureg_swizzle(t->pointSizeConst, 1,1,1,1)); + ureg_MIN( t->ureg, ureg_writemask(t->pointSizeResult, WRITEMASK_X), + ureg_src(t->outputs[t->pointSizeOutIndex]), + ureg_swizzle(t->pointSizeConst, 2,2,2,2)); + } + t->prevInstWrotePointSize = GL_FALSE; } /* Fix up all emitted labels: @@ -1019,7 +1223,7 @@ out: /** - * Tokens cannot be free with _mesa_free otherwise the builtin gallium + * Tokens cannot be free with free otherwise the builtin gallium * malloc debugging will get confused. */ void