diff options
Diffstat (limited to 'src/gallium/auxiliary/draw')
-rw-r--r-- | src/gallium/auxiliary/draw/SConscript | 3 | ||||
-rw-r--r-- | src/gallium/auxiliary/draw/draw_aapoint.c | 202 | ||||
-rw-r--r-- | src/gallium/auxiliary/draw/draw_context.c | 24 | ||||
-rw-r--r-- | src/gallium/auxiliary/draw/draw_pstipple.c | 4 | ||||
-rw-r--r-- | src/gallium/auxiliary/draw/draw_validate.c | 16 |
5 files changed, 138 insertions, 111 deletions
diff --git a/src/gallium/auxiliary/draw/SConscript b/src/gallium/auxiliary/draw/SConscript index 8e3a8caa74..3302dc44f7 100644 --- a/src/gallium/auxiliary/draw/SConscript +++ b/src/gallium/auxiliary/draw/SConscript @@ -3,6 +3,8 @@ Import('*') draw = env.ConvenienceLibrary( target = 'draw', source = [ + 'draw_aaline.c', + 'draw_aapoint.c', 'draw_clip.c', 'draw_vs_exec.c', 'draw_vs_sse.c', @@ -13,6 +15,7 @@ draw = env.ConvenienceLibrary( 'draw_flatshade.c', 'draw_offset.c', 'draw_prim.c', + 'draw_pstipple.c', 'draw_stipple.c', 'draw_twoside.c', 'draw_unfilled.c', diff --git a/src/gallium/auxiliary/draw/draw_aapoint.c b/src/gallium/auxiliary/draw/draw_aapoint.c index 43119cc70b..cae6fcd4d2 100644 --- a/src/gallium/auxiliary/draw/draw_aapoint.c +++ b/src/gallium/auxiliary/draw/draw_aapoint.c @@ -222,7 +222,7 @@ aa_transform_inst(struct tgsi_transform_context *ctx, * * Temp reg0 usage: * t0.x = distance of fragment from center point - * t0.y = boolean, is t0.x > 1 ? + * t0.y = boolean, is t0.x > 1.0, also misc temp usage * t0.z = temporary for computing 1/(1-k) value * t0.w = final coverage value */ @@ -313,9 +313,73 @@ aa_transform_inst(struct tgsi_transform_context *ctx, newInst.FullSrcRegisters[0].SrcRegister.Negate = 1; ctx->emit_instruction(ctx, &newInst); - /* SGT t0.y, t0.x, tex.z; # bool b = distance > k */ + + /* compute coverage factor = (1-d)/(1-k) */ + + /* SUB t0.z, tex.w, tex.z; # m = 1 - k */ newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_SGT; + newInst.Instruction.Opcode = TGSI_OPCODE_SUB; + newInst.Instruction.NumDstRegs = 1; + newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullDstRegisters[0].DstRegister.Index = tmp0; + newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z; + newInst.Instruction.NumSrcRegs = 2; + newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT; + newInst.FullSrcRegisters[0].SrcRegister.Index = texInput; + newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT; + newInst.FullSrcRegisters[1].SrcRegister.Index = texInput; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Z; + ctx->emit_instruction(ctx, &newInst); + + /* RCP t0.z, t0.z; # t0.z = 1 / m */ + newInst = tgsi_default_full_instruction(); + newInst.Instruction.Opcode = TGSI_OPCODE_RCP; + newInst.Instruction.NumDstRegs = 1; + newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullDstRegisters[0].DstRegister.Index = tmp0; + newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z; + newInst.Instruction.NumSrcRegs = 1; + newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0; + newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Z; + ctx->emit_instruction(ctx, &newInst); + + /* SUB t0.y, 1, t0.x; # d = 1 - d */ + newInst = tgsi_default_full_instruction(); + newInst.Instruction.Opcode = TGSI_OPCODE_SUB; + newInst.Instruction.NumDstRegs = 1; + newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullDstRegisters[0].DstRegister.Index = tmp0; + newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Y; + newInst.Instruction.NumSrcRegs = 2; + newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT; + newInst.FullSrcRegisters[0].SrcRegister.Index = texInput; + newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X; + ctx->emit_instruction(ctx, &newInst); + + /* MUL t0.w, t0.y, t0.z; # coverage = d * m */ + newInst = tgsi_default_full_instruction(); + newInst.Instruction.Opcode = TGSI_OPCODE_MUL; + newInst.Instruction.NumDstRegs = 1; + newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullDstRegisters[0].DstRegister.Index = tmp0; + newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W; + newInst.Instruction.NumSrcRegs = 2; + newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0; + newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y; + newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_Z; + ctx->emit_instruction(ctx, &newInst); + + /* SLE t0.y, t0.x, tex.z; # bool b = distance <= k */ + newInst = tgsi_default_full_instruction(); + newInst.Instruction.Opcode = TGSI_OPCODE_SLE; newInst.Instruction.NumDstRegs = 1; newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; newInst.FullDstRegisters[0].DstRegister.Index = tmp0; @@ -329,111 +393,40 @@ aa_transform_inst(struct tgsi_transform_context *ctx, newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_Z; ctx->emit_instruction(ctx, &newInst); - /* IF t0.y # if b then */ + /* CMP t0.w, -t0.y, tex.w, t0.w; + * # if -t0.y < 0 then + * t0.w = 1 + * else + * t0.w = t0.w + */ newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_IF; - newInst.Instruction.NumDstRegs = 0; - newInst.Instruction.NumSrcRegs = 1; + newInst.Instruction.Opcode = TGSI_OPCODE_CMP; + newInst.Instruction.NumDstRegs = 1; + newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullDstRegisters[0].DstRegister.Index = tmp0; + newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W; + newInst.Instruction.NumSrcRegs = 3; newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY; newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0; newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y; newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y; newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y; newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y; + newInst.FullSrcRegisters[0].SrcRegister.Negate = 1; + newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT; + newInst.FullSrcRegisters[1].SrcRegister.Index = texInput; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[2].SrcRegister.File = TGSI_FILE_TEMPORARY; + newInst.FullSrcRegisters[2].SrcRegister.Index = tmp0; + newInst.FullSrcRegisters[2].SrcRegister.SwizzleX = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[2].SrcRegister.SwizzleY = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[2].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W; + newInst.FullSrcRegisters[2].SrcRegister.SwizzleW = TGSI_SWIZZLE_W; ctx->emit_instruction(ctx, &newInst); - { - /* compute coverage factor = (1-d)/(1-k) */ - - /* SUB t0.z, tex.w, tex.z; # m = 1 - k */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_SUB; - newInst.Instruction.NumDstRegs = 1; - newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullDstRegisters[0].DstRegister.Index = tmp0; - newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z; - newInst.Instruction.NumSrcRegs = 2; - newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT; - newInst.FullSrcRegisters[0].SrcRegister.Index = texInput; - newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W; - newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT; - newInst.FullSrcRegisters[1].SrcRegister.Index = texInput; - newInst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Z; - ctx->emit_instruction(ctx, &newInst); - - /* RCP t0.z, t0.z; # t0.z = 1 / m */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_RCP; - newInst.Instruction.NumDstRegs = 1; - newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullDstRegisters[0].DstRegister.Index = tmp0; - newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z; - newInst.Instruction.NumSrcRegs = 1; - newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0; - newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Z; - ctx->emit_instruction(ctx, &newInst); - - /* SUB t0.x, 1, t0.x; # d = 1 - d */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_SUB; - newInst.Instruction.NumDstRegs = 1; - newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullDstRegisters[0].DstRegister.Index = tmp0; - newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X; - newInst.Instruction.NumSrcRegs = 2; - newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT; - newInst.FullSrcRegisters[0].SrcRegister.Index = texInput; - newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_W; - newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0; - newInst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_X; - ctx->emit_instruction(ctx, &newInst); - - /* MUL t0.w, t0.x, t0.z; # coverage = d * m */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_MUL; - newInst.Instruction.NumDstRegs = 1; - newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullDstRegisters[0].DstRegister.Index = tmp0; - newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W; - newInst.Instruction.NumSrcRegs = 2; - newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0; - newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_X; - newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0; - newInst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_Z; - ctx->emit_instruction(ctx, &newInst); - } - - /* ELSE */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_ELSE; - newInst.Instruction.NumDstRegs = 0; - newInst.Instruction.NumSrcRegs = 0; - ctx->emit_instruction(ctx, &newInst); - - { - /* MOV t0.w, tex.w; # coverage = 1.0 */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_MOV; - newInst.Instruction.NumDstRegs = 1; - newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY; - newInst.FullDstRegisters[0].DstRegister.Index = tmp0; - newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W; - newInst.Instruction.NumSrcRegs = 1; - newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT; - newInst.FullSrcRegisters[0].SrcRegister.Index = texInput; - ctx->emit_instruction(ctx, &newInst); - } - - /* ENDIF */ - newInst = tgsi_default_full_instruction(); - newInst.Instruction.Opcode = TGSI_OPCODE_ENDIF; - newInst.Instruction.NumDstRegs = 0; - newInst.Instruction.NumSrcRegs = 0; - ctx->emit_instruction(ctx, &newInst); } if (inst->Instruction.Opcode == TGSI_OPCODE_END) { @@ -516,7 +509,7 @@ generate_aapoint_fs(struct aapoint_stage *aapoint) (struct tgsi_token *) aapoint_fs.tokens, MAX, &transform.base); -#if 0 /* DEBUG */ +#if 1 /* DEBUG */ tgsi_dump(orig_fs->tokens, 0); tgsi_dump(aapoint_fs.tokens, 0); #endif @@ -613,13 +606,16 @@ aapoint_point(struct draw_stage *stage, struct prim_header *header) * ELSE * coverage = 1.0; // full coverage * ENDIF + * + * Note: the ELSEIF and ELSE clauses are actually implemented with CMP to + * avoid using IF/ELSE/ENDIF TGSI opcodes. */ #if !NORMALIZE - k = 1.0 / radius; - k = 1.0 - 2.0 * k + k * k; + k = 1.0f / radius; + k = 1.0f - 2.0f * k + k * k; #else - k = 1.0 - 1.0 / radius; + k = 1.0f - 1.0f / radius; #endif /* allocate/dup new verts */ diff --git a/src/gallium/auxiliary/draw/draw_context.c b/src/gallium/auxiliary/draw/draw_context.c index c28e78d33a..7dd1c6f6fa 100644 --- a/src/gallium/auxiliary/draw/draw_context.c +++ b/src/gallium/auxiliary/draw/draw_context.c @@ -244,14 +244,32 @@ draw_convert_wide_lines(struct draw_context *draw, boolean enable) /** - * The draw module may sometimes generate vertices with extra attributes - * (such as texcoords for AA lines). The driver can call this function - * to find those attributes. + * Ask the draw module for the location/slot of the given vertex attribute in + * a post-transformed vertex. + * + * With this function, drivers that use the draw module should have no reason + * to track the current vertex shader. + * + * Note that the draw module may sometimes generate vertices with extra + * attributes (such as texcoords for AA lines). The driver can call this + * function to find those attributes. + * + * Zero is returned if the attribute is not found since this is + * a don't care / undefined situtation. Returning -1 would be a bit more + * work for the drivers. */ int draw_find_vs_output(struct draw_context *draw, uint semantic_name, uint semantic_index) { + const struct pipe_shader_state *vs = draw->vertex_shader->state; + uint i; + for (i = 0; i < vs->num_outputs; i++) { + if (vs->output_semantic_name[i] == semantic_name && + vs->output_semantic_index[i] == semantic_index) + return i; + } + /* XXX there may be more than one extra vertex attrib. * For example, simulated gl_FragCoord and gl_PointCoord. */ diff --git a/src/gallium/auxiliary/draw/draw_pstipple.c b/src/gallium/auxiliary/draw/draw_pstipple.c index 4048abf856..1ab04cd959 100644 --- a/src/gallium/auxiliary/draw/draw_pstipple.c +++ b/src/gallium/auxiliary/draw/draw_pstipple.c @@ -133,7 +133,7 @@ pstip_transform_decl(struct tgsi_transform_context *ctx, pctx->maxSampler = (int) decl->u.DeclarationRange.Last; } else if (decl->Declaration.File == TGSI_FILE_INPUT) { - pctx->maxInput = MAX2(pctx->maxInput, decl->u.DeclarationRange.Last); + pctx->maxInput = MAX2(pctx->maxInput, (int) decl->u.DeclarationRange.Last); if (decl->Semantic.SemanticName == TGSI_SEMANTIC_POSITION) pctx->wincoordInput = (int) decl->u.DeclarationRange.First; } @@ -332,7 +332,7 @@ generate_pstip_fs(struct pstip_stage *pstip) if (transform.wincoordInput < 0) { pstip_fs.input_semantic_name[pstip_fs.num_inputs] = TGSI_SEMANTIC_POSITION; - pstip_fs.input_semantic_index[pstip_fs.num_inputs] = transform.maxInput; + pstip_fs.input_semantic_index[pstip_fs.num_inputs] = (ubyte)transform.maxInput; pstip_fs.num_inputs++; } diff --git a/src/gallium/auxiliary/draw/draw_validate.c b/src/gallium/auxiliary/draw/draw_validate.c index efd6793f2b..3a19dd4cd7 100644 --- a/src/gallium/auxiliary/draw/draw_validate.c +++ b/src/gallium/auxiliary/draw/draw_validate.c @@ -45,6 +45,7 @@ static struct draw_stage *validate_pipeline( struct draw_stage *stage ) struct draw_stage *next = draw->pipeline.rasterize; int need_det = 0; int precalc_flat = 0; + boolean wide_lines, wide_points; /* Set the validate's next stage to the rasterize stage, so that it * can be found later if needed for flushing. @@ -68,9 +69,18 @@ static struct draw_stage *validate_pipeline( struct draw_stage *stage ) next = draw->pipeline.aapoint; } - if ((draw->rasterizer->line_width != 1.0 && draw->convert_wide_lines - && !draw->rasterizer->line_smooth) || - (draw->rasterizer->point_size != 1.0 && draw->convert_wide_points) || + /* drawing wide lines? */ + wide_lines = (draw->rasterizer->line_width != 1.0 + && draw->convert_wide_lines + && !draw->rasterizer->line_smooth); + + /* drawing large points? */ + wide_points = (draw->rasterizer->point_size != 1.0 + && draw->convert_wide_points + && !draw->pipeline.aapoint); + + if (wide_lines || + wide_points || draw->rasterizer->point_sprite) { draw->pipeline.wide->next = next; next = draw->pipeline.wide; |