[Mesa-dev] [PATCH] r600/llvm: Store inputs in function arguments
Tom Stellard
tom at stellard.net
Mon Nov 4 09:02:00 PST 2013
On Sun, Nov 03, 2013 at 06:58:18PM +0100, Vincent Lejeune wrote:
> ---
> src/gallium/drivers/r600/r600_llvm.c | 125 ++++++++++++---------
> src/gallium/drivers/r600/r600_shader.c | 2 +
> src/gallium/drivers/radeon/radeon_llvm.h | 1 +
> .../drivers/radeon/radeon_setup_tgsi_llvm.c | 2 +-
> 4 files changed, 75 insertions(+), 55 deletions(-)
As discussed on IRC, I think this will need some ifdefs to remain
compatible with older LLVM. Plus, I have two comments below:
>
> diff --git a/src/gallium/drivers/r600/r600_llvm.c b/src/gallium/drivers/r600/r600_llvm.c
> index 5afe3cb..8dcda1a 100644
> --- a/src/gallium/drivers/r600/r600_llvm.c
> +++ b/src/gallium/drivers/r600/r600_llvm.c
> @@ -87,37 +87,50 @@ static void llvm_load_system_value(
> }
>
> static LLVMValueRef
> -llvm_load_input_helper(
> - struct radeon_llvm_context * ctx,
> - unsigned idx, int interp, int ij_index)
> +llvm_load_input_vector(
> + struct radeon_llvm_context * ctx, unsigned location, unsigned ijregs,
> + boolean interp)
> {
> - const struct lp_build_context * bb = &ctx->soa.bld_base.base;
> - LLVMValueRef arg[2];
> - int arg_count;
> - const char * intrinsic;
> -
> - arg[0] = lp_build_const_int32(bb->gallivm, idx);
> -
> - if (interp) {
> - intrinsic = "llvm.R600.interp.input";
> - arg[1] = lp_build_const_int32(bb->gallivm, ij_index);
> - arg_count = 2;
> - } else {
> - intrinsic = "llvm.R600.load.input";
> - arg_count = 1;
> - }
> -
> - return build_intrinsic(bb->gallivm->builder, intrinsic,
> - bb->elem_type, &arg[0], arg_count, LLVMReadNoneAttribute);
> + LLVMTypeRef VecType;
> + LLVMValueRef Args[2] = {
> + lp_build_const_int32(&(ctx->gallivm), location)
> + };
> + unsigned ArgCount = 1;
> + if (interp) {
> + VecType = LLVMVectorType(ctx->soa.bld_base.base.elem_type, 2);
> + LLVMValueRef IJIndex = LLVMGetParam(ctx->main_fn, ijregs / 2);
> + Args[ArgCount++] = LLVMBuildExtractElement(ctx->gallivm.builder, IJIndex,
> + lp_build_const_int32(&(ctx->gallivm), 2 * (ijregs % 2)), "");
> + Args[ArgCount++] = LLVMBuildExtractElement(ctx->gallivm.builder, IJIndex,
> + lp_build_const_int32(&(ctx->gallivm), 2 * (ijregs % 2) + 1), "");
> + LLVMValueRef HalfVec[2] = {
> + build_intrinsic(ctx->gallivm.builder, "llvm.R600.interp.xy",
> + VecType, Args, ArgCount, LLVMReadNoneAttribute),
> + build_intrinsic(ctx->gallivm.builder, "llvm.R600.interp.zw",
> + VecType, Args, ArgCount, LLVMReadNoneAttribute)
> + };
> + LLVMValueRef MaskInputs[4] = {
> + lp_build_const_int32(&(ctx->gallivm), 0),
> + lp_build_const_int32(&(ctx->gallivm), 1),
> + lp_build_const_int32(&(ctx->gallivm), 2),
> + lp_build_const_int32(&(ctx->gallivm), 3)
> + };
> + LLVMValueRef Mask = LLVMConstVector(MaskInputs, 4);
> + return LLVMBuildShuffleVector(ctx->gallivm.builder, HalfVec[0], HalfVec[1],
> + Mask, "");
> + } else {
> + VecType = LLVMVectorType(ctx->soa.bld_base.base.elem_type, 4);
> + return build_intrinsic(ctx->gallivm.builder, "llvm.R600.interp.const",
> + VecType, Args, ArgCount, LLVMReadNoneAttribute);
> + }
> }
>
> static LLVMValueRef
> llvm_face_select_helper(
> struct radeon_llvm_context * ctx,
> - unsigned face_loc, LLVMValueRef front_color, LLVMValueRef back_color)
> + LLVMValueRef face, LLVMValueRef front_color, LLVMValueRef back_color)
> {
> const struct lp_build_context * bb = &ctx->soa.bld_base.base;
> - LLVMValueRef face = llvm_load_input_helper(ctx, face_loc, 0, 0);
> LLVMValueRef is_front = LLVMBuildFCmp(
> bb->gallivm->builder, LLVMRealUGT, face,
> lp_build_const_float(bb->gallivm, 0.0f), "");
> @@ -132,50 +145,46 @@ static void llvm_load_input(
> {
> const struct r600_shader_io * input = &ctx->r600_inputs[input_index];
> unsigned chan;
> - unsigned interp = 0;
> - int ij_index;
> int two_side = (ctx->two_side && input->name == TGSI_SEMANTIC_COLOR);
> LLVMValueRef v;
> + boolean require_interp_intrinsic = ctx->chip_class >= EVERGREEN &&
> + ctx->type == TGSI_PROCESSOR_FRAGMENT;
>
> - if (ctx->chip_class >= EVERGREEN && ctx->type == TGSI_PROCESSOR_FRAGMENT &&
> - input->spi_sid) {
> - interp = 1;
> - ij_index = (input->interpolate > 0) ? input->ij_index : -1;
> - }
> + if (require_interp_intrinsic && input->spi_sid) {
> + v = llvm_load_input_vector(ctx, input->lds_pos, input->ij_index,
> + (input->interpolate > 0));
> + } else
> + v = LLVMGetParam(ctx->main_fn, input->gpr);
>
> - for (chan = 0; chan < 4; chan++) {
> - unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
> - int loc;
> + if (two_side) {
> + struct r600_shader_io * back_input =
> + &ctx->r600_inputs[input->back_color_input];
> + LLVMValueRef v2;
> + LLVMValueRef face = LLVMGetParam(ctx->main_fn, ctx->face_gpr);
> + face = LLVMBuildExtractElement(ctx->gallivm.builder, face,
> + lp_build_const_int32(&(ctx->gallivm), 0), "");
>
> - if (interp) {
> - loc = 4 * input->lds_pos + chan;
> - } else {
> - if (input->name == TGSI_SEMANTIC_FACE)
> - loc = 4 * ctx->face_gpr;
> + if (require_interp_intrinsic && back_input->spi_sid)
> + v2 = llvm_load_input_vector(ctx, back_input->lds_pos,
> + back_input->ij_index, (back_input->interpolate > 0));
> else
> - loc = 4 * input->gpr + chan;
> + v2 = LLVMGetParam(ctx->main_fn, back_input->gpr);
> + v = llvm_face_select_helper(ctx, face, v, v2);
> }
>
> - v = llvm_load_input_helper(ctx, loc, interp, ij_index);
> + for (chan = 0; chan < 4; chan++) {
> + unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
>
> - if (two_side) {
> - struct r600_shader_io * back_input =
> - &ctx->r600_inputs[input->back_color_input];
> - int back_loc = interp ? back_input->lds_pos : back_input->gpr;
> - LLVMValueRef v2;
> + ctx->inputs[soa_index] = LLVMBuildExtractElement(ctx->gallivm.builder, v,
> + lp_build_const_int32(&(ctx->gallivm), chan), "");
>
> - back_loc = 4 * back_loc + chan;
> - v2 = llvm_load_input_helper(ctx, back_loc, interp, ij_index);
> - v = llvm_face_select_helper(ctx, 4 * ctx->face_gpr, v, v2);
> - } else if (input->name == TGSI_SEMANTIC_POSITION &&
> + if (input->name == TGSI_SEMANTIC_POSITION &&
> ctx->type == TGSI_PROCESSOR_FRAGMENT && chan == 3) {
> /* RCP for fragcoord.w */
> - v = LLVMBuildFDiv(ctx->gallivm.builder,
> + ctx->inputs[soa_index] = LLVMBuildFDiv(ctx->gallivm.builder,
> lp_build_const_float(&(ctx->gallivm), 1.0f),
> - v, "");
> + ctx->inputs[soa_index], "");
> }
> -
> - ctx->inputs[soa_index] = v;
> }
> }
>
> @@ -657,7 +666,15 @@ LLVMModuleRef r600_tgsi_llvm(
> struct tgsi_shader_info shader_info;
> struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
> radeon_llvm_context_init(ctx);
> - radeon_llvm_create_func(ctx, NULL, 0);
> + LLVMTypeRef Arguments[32];
> + unsigned ArgumentsCount = 0;
> + for (unsigned i = 0; i < ctx->inputs_count; i++)
> + Arguments[ArgumentsCount++] = LLVMVectorType(bld_base->base.elem_type, 4);
> + radeon_llvm_create_func(ctx, Arguments, ArgumentsCount);
> + for (unsigned i = 0; i < ctx->inputs_count; i++) {
> + LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);
> + LLVMAddAttribute(P, LLVMInRegAttribute);
> + }
> tgsi_scan_shader(tokens, &shader_info);
>
> bld_base->info = &shader_info;
> diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
> index aed2100..ae134d1 100644
> --- a/src/gallium/drivers/r600/r600_shader.c
> +++ b/src/gallium/drivers/r600/r600_shader.c
> @@ -899,6 +899,7 @@ static int r600_shader_from_tgsi(struct r600_screen *rscreen,
> #ifdef R600_USE_LLVM
> use_llvm = !(rscreen->b.debug_flags & DBG_NO_LLVM);
> #endif
> +
Extra whitespace here.
> ctx.bc = &shader->bc;
> ctx.shader = shader;
> ctx.native_integers = true;
> @@ -1102,6 +1103,7 @@ static int r600_shader_from_tgsi(struct r600_screen *rscreen,
> radeon_llvm_ctx.type = ctx.type;
> radeon_llvm_ctx.two_side = shader->two_side;
> radeon_llvm_ctx.face_gpr = ctx.face_gpr;
> + radeon_llvm_ctx.inputs_count = ctx.shader->ninput + 1;
> radeon_llvm_ctx.r600_inputs = ctx.shader->input;
> radeon_llvm_ctx.r600_outputs = ctx.shader->output;
> radeon_llvm_ctx.color_buffer_count = max_color_exports;
> diff --git a/src/gallium/drivers/radeon/radeon_llvm.h b/src/gallium/drivers/radeon/radeon_llvm.h
> index ef09dc8..2cab6b0 100644
> --- a/src/gallium/drivers/radeon/radeon_llvm.h
> +++ b/src/gallium/drivers/radeon/radeon_llvm.h
> @@ -60,6 +60,7 @@ struct radeon_llvm_context {
> unsigned face_gpr;
> unsigned two_side;
> unsigned clip_vertex;
> + unsigned inputs_count;
> struct r600_shader_io * r600_inputs;
> struct r600_shader_io * r600_outputs;
> struct pipe_stream_output_info *stream_outputs;
> diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> index 43ebe7f..9ed671c 100644
> --- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> +++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> @@ -1252,7 +1252,7 @@ void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
> bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
> bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
> bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.AMDIL.exp.";
> + bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.exp2.f32";
This change seems unrelated.
> bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_readonly;
> bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "floor";
> bld_base->op_actions[TGSI_OPCODE_FRC].emit = build_tgsi_intrinsic_nomem;
> --
> 1.8.3.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list