From f79ef95df4f19124c24e59583bf9fb1e347d8f2b Mon Sep 17 00:00:00 2001 From: Maciej Cencora Date: Sun, 5 Jul 2009 02:32:51 +0200 Subject: [PATCH] r300: rewrite FOGC and HPOS attribs handling Rewrite vertex and fragment programs so that we don't have to do any hacks on lower level. --- src/mesa/drivers/dri/r300/r300_context.h | 8 +- src/mesa/drivers/dri/r300/r300_emit.c | 9 --- .../drivers/dri/r300/r300_fragprog_common.c | 62 ++++++++++++++- src/mesa/drivers/dri/r300/r300_shader.c | 3 +- src/mesa/drivers/dri/r300/r300_state.c | 53 ------------- src/mesa/drivers/dri/r300/r300_swtcl.c | 38 ++++----- src/mesa/drivers/dri/r300/r300_vertprog.c | 79 +++++++++++-------- 7 files changed, 130 insertions(+), 122 deletions(-) diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h index 6f99e490500..32104b8158b 100644 --- a/src/mesa/drivers/dri/r300/r300_context.h +++ b/src/mesa/drivers/dri/r300/r300_context.h @@ -410,6 +410,8 @@ struct r300_vertex_program { struct r300_vertex_program_key { GLuint FpReads; + GLuint FogAttr; + GLuint WPosAttr; } key; struct r300_vertex_shader_hw_code { @@ -425,7 +427,6 @@ struct r300_vertex_program { int pos_end; int num_temporaries; /* Number of temp vars used by program */ - int wpos_idx; int inputs[VERT_ATTRIB_MAX]; int outputs[VERT_RESULT_MAX]; }; @@ -560,6 +561,11 @@ struct r300_fragment_program { GLuint optimization; struct r300_fragment_program *next; + + /* attribute that we are sending the WPOS in */ + gl_frag_attrib wpos_attr; + /* attribute that we are sending the fog coordinate in */ + gl_frag_attrib fog_attr; }; struct r300_fragment_program_cont { diff --git a/src/mesa/drivers/dri/r300/r300_emit.c b/src/mesa/drivers/dri/r300/r300_emit.c index c3817721dc4..707d1284edb 100644 --- a/src/mesa/drivers/dri/r300/r300_emit.c +++ b/src/mesa/drivers/dri/r300/r300_emit.c @@ -116,15 +116,6 @@ GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads) } } - if (fp_reads & FRAG_BIT_WPOS) { - ret |= (4 << (3 * first_free_texcoord)); - ++first_free_texcoord; - } - - if (vp_writes & (1 << VERT_RESULT_FOGC) && fp_reads & FRAG_BIT_FOGC) { - ret |= 4 << (3 * first_free_texcoord); - } - if (first_free_texcoord > 8) { fprintf(stderr, "\tout of free texcoords\n"); _mesa_exit(-1); diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_common.c b/src/mesa/drivers/dri/r300/r300_fragprog_common.c index b25cf24007a..e90be9b7f85 100644 --- a/src/mesa/drivers/dri/r300/r300_fragprog_common.c +++ b/src/mesa/drivers/dri/r300/r300_fragprog_common.c @@ -69,8 +69,10 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler) { GLuint InputsRead = compiler->fp->Base->InputsRead; - if (!(InputsRead & FRAG_BIT_WPOS)) + if (!(InputsRead & FRAG_BIT_WPOS)) { + compiler->fp->wpos_attr = FRAG_ATTRIB_MAX; return; + } static gl_state_index tokens[STATE_LENGTH] = { STATE_INTERNAL, STATE_R300_WINDOW_DIMENSION, 0, 0, 0 @@ -78,10 +80,23 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler) struct prog_instruction *fpi; GLuint window_index; int i = 0; + + for (i = FRAG_ATTRIB_TEX0; i <= FRAG_ATTRIB_TEX7; ++i) + { + if (!(InputsRead & (1 << i))) { + InputsRead &= ~(1 << FRAG_ATTRIB_WPOS); + InputsRead |= 1 << i; + compiler->fp->Base->InputsRead = InputsRead; + compiler->fp->wpos_attr = i; + break; + } + } + GLuint tempregi = _mesa_find_free_register(compiler->program, PROGRAM_TEMPORARY); _mesa_insert_instructions(compiler->program, 0, 3); fpi = compiler->program->Instructions; + i = 0; /* perspective divide */ fpi[i].Opcode = OPCODE_RCP; @@ -92,7 +107,7 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler) fpi[i].DstReg.CondMask = COND_TR; fpi[i].SrcReg[0].File = PROGRAM_INPUT; - fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS; + fpi[i].SrcReg[0].Index = compiler->fp->wpos_attr; fpi[i].SrcReg[0].Swizzle = SWIZZLE_WWWW; i++; @@ -104,7 +119,7 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler) fpi[i].DstReg.CondMask = COND_TR; fpi[i].SrcReg[0].File = PROGRAM_INPUT; - fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS; + fpi[i].SrcReg[0].Index = compiler->fp->wpos_attr; fpi[i].SrcReg[0].Swizzle = SWIZZLE_XYZW; fpi[i].SrcReg[1].File = PROGRAM_TEMPORARY; @@ -147,6 +162,45 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler) } } +static void rewriteFog(struct r300_fragment_program_compiler *compiler) +{ + struct r300_fragment_program *fp = compiler->fp; + GLuint InputsRead; + int i; + + InputsRead = fp->Base->InputsRead; + + if (!(InputsRead & FRAG_BIT_FOGC)) { + fp->fog_attr = FRAG_ATTRIB_MAX; + return; + } + + for (i = FRAG_ATTRIB_TEX0; i <= FRAG_ATTRIB_TEX7; ++i) + { + if (!(InputsRead & (1 << i))) { + InputsRead &= ~(1 << FRAG_ATTRIB_FOGC); + InputsRead |= 1 << i; + fp->Base->InputsRead = InputsRead; + fp->fog_attr = i; + break; + } + } + + { + struct prog_instruction *inst; + + inst = compiler->program->Instructions; + while (inst->Opcode != OPCODE_END) { + const int src_regs = _mesa_num_inst_src_regs(inst->Opcode); + for (i = 0; i < src_regs; ++i) { + if (inst->SrcReg[i].File == PROGRAM_INPUT && inst->SrcReg[i].Index == FRAG_ATTRIB_FOGC) + inst->SrcReg[i].Index = fp->fog_attr; + } + ++inst; + } + } +} + static GLuint build_dtm(GLuint depthmode) { switch(depthmode) { @@ -204,6 +258,8 @@ void r300TranslateFragmentShader(GLcontext *ctx, struct r300_fragment_program *f insert_WPOS_trailer(&compiler); + rewriteFog(&compiler); + if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) { struct radeon_program_transformation transformations[] = { { &r500_transform_TEX, &compiler }, diff --git a/src/mesa/drivers/dri/r300/r300_shader.c b/src/mesa/drivers/dri/r300/r300_shader.c index 854eb5d80ae..62228a3786e 100644 --- a/src/mesa/drivers/dri/r300/r300_shader.c +++ b/src/mesa/drivers/dri/r300/r300_shader.c @@ -66,8 +66,7 @@ static struct gl_program *r300NewProgram(GLcontext * ctx, GLenum target, case GL_VERTEX_STATE_PROGRAM_NV: case GL_VERTEX_PROGRAM_ARB: vp = CALLOC_STRUCT(r300_vertex_program_cont); - return _mesa_init_vertex_program(ctx, &vp->mesa_program, - target, id); + return _mesa_init_vertex_program(ctx, &vp->mesa_program, target, id); case GL_FRAGMENT_PROGRAM_NV: case GL_FRAGMENT_PROGRAM_ARB: diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c index 0f3198e792d..bae2f0f3cb0 100644 --- a/src/mesa/drivers/dri/r300/r300_state.c +++ b/src/mesa/drivers/dri/r300/r300_state.c @@ -1519,29 +1519,6 @@ static void r300SetupRSUnit(GLcontext * ctx) ++fp_reg; } - if (InputsRead & FRAG_BIT_WPOS) { - r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_S(0) | R300_RS_SEL_T(1) | R300_RS_SEL_R(2) | R300_RS_SEL_Q(3) | R300_RS_TEX_PTR(rs_tex_count); - r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R300_RS_INST_TEX_ID(tex_ip) | R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_reg); - InputsRead &= ~FRAG_BIT_WPOS; - rs_tex_count += 4; - ++tex_ip; - ++fp_reg; - } - - if (InputsRead & FRAG_BIT_FOGC) { - if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_FOGC, _TNL_ATTRIB_FOG)) { - r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_S(0) | R300_RS_SEL_T(R300_RS_SEL_K0) | R300_RS_SEL_R(R300_RS_SEL_K0); - r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_Q(R300_RS_SEL_K1) | R300_RS_TEX_PTR(rs_tex_count); - r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R300_RS_INST_TEX_ID(tex_ip) | R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_reg); - InputsRead &= ~FRAG_BIT_FOGC; - rs_tex_count += 4; - ++tex_ip; - ++fp_reg; - } else { - WARN_ONCE("fragprog wants fogc, vp doesn't provide it\n"); - } - } - /* Setup default color if no color or tex was set */ if (rs_tex_count == 0 && col_ip == 0) { r300->hw.rr.cmd[R300_RR_INST_0] = R300_RS_INST_COL_ID(0) | R300_RS_INST_COL_ADDR(0); @@ -1640,36 +1617,6 @@ static void r500SetupRSUnit(GLcontext * ctx) ++fp_reg; } - if (InputsRead & FRAG_BIT_WPOS) { - r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= ((rs_tex_count + 0) << R500_RS_IP_TEX_PTR_S_SHIFT) | - ((rs_tex_count + 1) << R500_RS_IP_TEX_PTR_T_SHIFT) | - ((rs_tex_count + 2) << R500_RS_IP_TEX_PTR_R_SHIFT) | - ((rs_tex_count + 3) << R500_RS_IP_TEX_PTR_Q_SHIFT); - - r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R500_RS_INST_TEX_ID(tex_ip) | R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_reg); - InputsRead &= ~FRAG_BIT_WPOS; - rs_tex_count += 4; - ++tex_ip; - ++fp_reg; - } - - if (InputsRead & FRAG_BIT_FOGC) { - if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_FOGC, _TNL_ATTRIB_FOG)) { - r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= (rs_tex_count << R500_RS_IP_TEX_PTR_S_SHIFT) | - (R500_RS_IP_PTR_K0 << R500_RS_IP_TEX_PTR_T_SHIFT) | - (R500_RS_IP_PTR_K0 << R500_RS_IP_TEX_PTR_R_SHIFT) | - (R500_RS_IP_PTR_K1 << R500_RS_IP_TEX_PTR_Q_SHIFT); - - r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R500_RS_INST_TEX_ID(tex_ip) | R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_reg); - InputsRead &= ~FRAG_BIT_FOGC; - rs_tex_count += 4; - ++tex_ip; - ++fp_reg; - } else { - WARN_ONCE("fragprog wants fogc, vp doesn't provide it\n"); - } - } - /* Setup default color if no color or tex was set */ if (rs_tex_count == 0 && col_ip == 0) { r300->hw.rr.cmd[R300_RR_INST_0] = R500_RS_INST_COL_ID(0) | R500_RS_INST_COL_ADDR(0); diff --git a/src/mesa/drivers/dri/r300/r300_swtcl.c b/src/mesa/drivers/dri/r300/r300_swtcl.c index db4ccce6f1f..d73508d36e3 100644 --- a/src/mesa/drivers/dri/r300/r300_swtcl.c +++ b/src/mesa/drivers/dri/r300/r300_swtcl.c @@ -192,31 +192,23 @@ void r300ChooseSwtclVertexFormat(GLcontext *ctx, GLuint *_InputsRead, GLuint *_ } } - /* RS can't put fragment position on the pixel stack, so stuff it in texcoord if needed */ - if (fp_reads & FRAG_BIT_WPOS) { - if (first_free_tex >= ctx->Const.MaxTextureUnits) { - fprintf(stderr, "\tout of free texcoords to write w pos\n"); - _mesa_exit(-1); - } + if (rmesa->selected_fp->wpos_attr != FRAG_ATTRIB_MAX) { + int tex_id = rmesa->selected_fp->wpos_attr - FRAG_ATTRIB_TEX0; - InputsRead |= 1 << (VERT_ATTRIB_TEX0 + first_free_tex); - OutputsWritten |= 1 << (VERT_RESULT_TEX0 + first_free_tex); - EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_4F ); - ADD_ATTR(VERT_ATTRIB_POS, R300_DATA_TYPE_FLOAT_4, SWTCL_OVM_TEX(first_free_tex), SWIZZLE_XYZW, MASK_XYZW, 0); - ++first_free_tex; + VB->AttribPtr[VERT_ATTRIB_TEX0 + tex_id] = VB->AttribPtr[VERT_ATTRIB_POS]; + RENDERINPUTS_SET(tnl->render_inputs_bitset, _TNL_ATTRIB_TEX0 + tex_id); } - if (fp_reads & FRAG_BIT_FOGC) { - if (first_free_tex >= ctx->Const.MaxTextureUnits) { - fprintf(stderr, "\tout of free texcoords to write fog coordinate\n"); - _mesa_exit(-1); - } + if (rmesa->selected_fp->fog_attr != FRAG_ATTRIB_MAX) { + int tex_id = rmesa->selected_fp->fog_attr - FRAG_ATTRIB_TEX0; - InputsRead |= 1 << VERT_ATTRIB_FOG; - OutputsWritten |= 1 << VERT_RESULT_FOGC; - GLuint swiz = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ZERO); - EMIT_ATTR( _TNL_ATTRIB_FOG, EMIT_1F ); - ADD_ATTR(VERT_ATTRIB_FOG, R300_DATA_TYPE_FLOAT_1, SWTCL_OVM_TEX(first_free_tex), swiz, MASK_XYZW, 0); + VB->AttribPtr[VERT_ATTRIB_TEX0 + tex_id] = VB->AttribPtr[VERT_ATTRIB_FOG]; + RENDERINPUTS_SET(tnl->render_inputs_bitset, _TNL_ATTRIB_TEX0 + tex_id); + } + + if (first_free_tex >= ctx->Const.MaxTextureUnits) { + fprintf(stderr, "\tout of free texcoords to write fog coordinate\n"); + _mesa_exit(-1); } R300_NEWPRIM(rmesa); @@ -497,11 +489,13 @@ void r300RenderStart(GLcontext *ctx) r300ContextPtr rmesa = R300_CONTEXT( ctx ); r300ChooseRenderState(ctx); + + r300UpdateShaders(rmesa); + r300PrepareVertices(ctx); r300ValidateBuffers(ctx); - r300UpdateShaders(rmesa); r300UpdateShaderStates(rmesa); r300EmitCacheFlush(rmesa); diff --git a/src/mesa/drivers/dri/r300/r300_vertprog.c b/src/mesa/drivers/dri/r300/r300_vertprog.c index 0b048303508..6f9b4db40f7 100644 --- a/src/mesa/drivers/dri/r300/r300_vertprog.c +++ b/src/mesa/drivers/dri/r300/r300_vertprog.c @@ -1223,8 +1223,7 @@ void r300TranslateVertexShader(struct r300_vertex_program *vp) } } -static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog, - GLuint temp_index) +static void insert_wpos(struct gl_program *prog, GLuint temp_index, int tex_id) { struct prog_instruction *vpi; @@ -1248,7 +1247,7 @@ static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog, vpi->Opcode = OPCODE_MOV; vpi->DstReg.File = PROGRAM_OUTPUT; - vpi->DstReg.Index = VERT_RESULT_TEX0 + vp->wpos_idx; + vpi->DstReg.Index = VERT_RESULT_TEX0 + tex_id; vpi->DstReg.WriteMask = WRITEMASK_XYZW; vpi->DstReg.CondMask = COND_TR; @@ -1259,12 +1258,9 @@ static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog, ++vpi; vpi->Opcode = OPCODE_END; - - prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + vp->wpos_idx); } -static void pos_as_texcoord(struct r300_vertex_program *vp, - struct gl_program *prog) +static void pos_as_texcoord(struct gl_program *prog, int tex_id) { struct prog_instruction *vpi; GLuint tempregi = prog->NumTemporaries; @@ -1278,7 +1274,37 @@ static void pos_as_texcoord(struct r300_vertex_program *vp, } } - insert_wpos(vp, prog, tempregi); + insert_wpos(prog, tempregi, tex_id); + + prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + tex_id); +} + +/** + @TODO + We can put X001 swizzle only if input components are directly mapped from output components. + For some insts we need to skip source swizzles and add: MOV OUTPUT[fog_attr].yzw, CONST[0].0001 + */ +static void fog_as_texcoord(struct gl_program *prog, int tex_id) +{ + struct prog_instruction *vpi; + int i; + + vpi = prog->Instructions; + while (vpi->Opcode != OPCODE_END) { + if (vpi->DstReg.File == PROGRAM_OUTPUT && vpi->DstReg.Index == VERT_RESULT_FOGC) { + vpi->DstReg.Index = VERT_RESULT_TEX0 + tex_id; + vpi->DstReg.WriteMask = WRITEMASK_XYZW; + + for (i = 0; i < _mesa_num_inst_src_regs(vpi->Opcode); ++i) { + vpi->SrcReg[i].Swizzle = combineSwizzles(vpi->SrcReg[i].Swizzle, SWIZZLE_X, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ONE); + } + } + + ++vpi; + } + + prog->OutputsWritten &= ~(1 << VERT_RESULT_FOGC); + prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + tex_id); } static int translateABS(struct gl_program *prog, int pos) @@ -1513,16 +1539,15 @@ static void addArtificialOutputs(GLcontext *ctx, struct gl_program *prog) static struct r300_vertex_program *build_program(GLcontext *ctx, struct r300_vertex_program_key *wanted_key, - const struct gl_vertex_program *mesa_vp, - GLint wpos_idx) + const struct gl_vertex_program *mesa_vp) { + r300ContextPtr r300 = R300_CONTEXT(ctx); struct r300_vertex_program *vp; struct gl_program *prog; vp = _mesa_calloc(sizeof(*vp)); vp->Base = (struct gl_vertex_program *) _mesa_clone_program(ctx, &mesa_vp->Base); _mesa_memcpy(&vp->key, wanted_key, sizeof(vp->key)); - vp->wpos_idx = wpos_idx; prog = &vp->Base->Base; @@ -1532,12 +1557,16 @@ static struct r300_vertex_program *build_program(GLcontext *ctx, fflush(stdout); } - if (mesa_vp->IsPositionInvariant) { + if (vp->Base->IsPositionInvariant) { _mesa_insert_mvp_code(ctx, vp->Base); } - if (wpos_idx > -1) { - pos_as_texcoord(vp, prog); + if (r300->selected_fp->wpos_attr != FRAG_ATTRIB_MAX) { + pos_as_texcoord(&vp->Base->Base, r300->selected_fp->wpos_attr - FRAG_ATTRIB_TEX0); + } + + if (r300->selected_fp->fog_attr != FRAG_ATTRIB_MAX) { + fog_as_texcoord(&vp->Base->Base, r300->selected_fp->fog_attr - FRAG_ATTRIB_TEX0); } addArtificialOutputs(ctx, prog); @@ -1562,34 +1591,20 @@ struct r300_vertex_program * r300SelectVertexShader(GLcontext *ctx) struct r300_vertex_program_key wanted_key = { 0 }; struct r300_vertex_program_cont *vpc; struct r300_vertex_program *vp; - GLint wpos_idx; vpc = (struct r300_vertex_program_cont *)ctx->VertexProgram._Current; wanted_key.FpReads = r300->selected_fp->Base->InputsRead; + wanted_key.FogAttr = r300->selected_fp->fog_attr; + wanted_key.WPosAttr = r300->selected_fp->wpos_attr; - for (vp = vpc->progs; vp; vp = vp->next) + for (vp = vpc->progs; vp; vp = vp->next) { if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key)) == 0) { return r300->selected_vp = vp; } - - wpos_idx = -1; - if (wanted_key.FpReads & FRAG_BIT_WPOS) { - GLint i; - - for (i = 0; i < ctx->Const.MaxTextureUnits; i++) - if (!(wanted_key.FpReads & (FRAG_BIT_TEX(i)))) - break; - - if (i == ctx->Const.MaxTextureUnits) { - fprintf(stderr, "\tno free texcoord found\n"); - _mesa_exit(-1); - } - - wpos_idx = i; } - vp = build_program(ctx, &wanted_key, &vpc->mesa_program, wpos_idx); + vp = build_program(ctx, &wanted_key, &vpc->mesa_program); vp->next = vpc->progs; vpc->progs = vp; -- 2.30.2