X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fradeonsi%2Fsi_shader.c;h=2e4cf0884ca5e9b0c0adfb9bf0d6035c7d484112;hb=37916a66b1ab8943e052874632785f033f2f04e2;hp=f727e2ec6a43b19baa291cc4a00363ad71a3dc81;hpb=1e3aab4cd08af580f65ee5adaebd8c370bda1fde;p=mesa.git diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index f727e2ec6a4..2e4cf0884ca 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -25,14 +25,9 @@ #include #include "util/u_memory.h" -#include "util/u_string.h" -#include "tgsi/tgsi_build.h" #include "tgsi/tgsi_strings.h" -#include "tgsi/tgsi_util.h" -#include "tgsi/tgsi_dump.h" #include "tgsi/tgsi_from_mesa.h" -#include "ac_binary.h" #include "ac_exp_param.h" #include "ac_shader_util.h" #include "ac_rtld.h" @@ -50,15 +45,7 @@ static const char scratch_rsrc_dword0_symbol[] = static const char scratch_rsrc_dword1_symbol[] = "SCRATCH_RSRC_DWORD1"; -static void si_init_shader_ctx(struct si_shader_context *ctx, - struct si_screen *sscreen, - struct ac_llvm_compiler *compiler, - unsigned wave_size, - bool nir); - -static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data); +static void si_llvm_emit_barrier(struct si_shader_context *ctx); static void si_dump_shader_key(const struct si_shader *shader, FILE *f); @@ -285,7 +272,7 @@ static LLVMValueRef get_tcs_out_patch_stride(struct si_shader_context *ctx) if (ctx->shader->key.mono.u.ff_tcs_inputs_to_copy) return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 0, 13); - const struct tgsi_shader_info *info = &ctx->shader->selector->info; + const struct si_shader_info *info = &ctx->shader->selector->info; unsigned tcs_out_vertices = info->properties[TGSI_PROPERTY_TCS_VERTICES_OUT]; unsigned vertex_dw_stride = get_tcs_out_vertex_dw_stride_constant(ctx); unsigned num_patch_outputs = util_last_bit64(ctx->shader->selector->patch_outputs_written); @@ -396,7 +383,7 @@ void si_llvm_load_input_vs( unsigned input_index, LLVMValueRef out[4]) { - const struct tgsi_shader_info *info = &ctx->shader->selector->info; + const struct si_shader_info *info = &ctx->shader->selector->info; unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]; if (vs_blit_property) { @@ -468,19 +455,20 @@ void si_llvm_load_input_vs( return; } + unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs; union si_vs_fix_fetch fix_fetch; - LLVMValueRef t_list_ptr; - LLVMValueRef t_offset; - LLVMValueRef t_list; + LLVMValueRef vb_desc; LLVMValueRef vertex_index; LLVMValueRef tmp; - /* Load the T list */ - t_list_ptr = ac_get_arg(&ctx->ac, ctx->vertex_buffers); - - t_offset = LLVMConstInt(ctx->i32, input_index, 0); - - t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset); + if (input_index < num_vbos_in_user_sgprs) { + vb_desc = ac_get_arg(&ctx->ac, ctx->vb_descriptors[input_index]); + } else { + unsigned index= input_index - num_vbos_in_user_sgprs; + vb_desc = ac_build_load_to_sgpr(&ctx->ac, + ac_get_arg(&ctx->ac, ctx->vertex_buffers), + LLVMConstInt(ctx->i32, index, 0)); + } vertex_index = LLVMGetParam(ctx->main_fn, ctx->vertex_index0.arg_index + @@ -501,7 +489,7 @@ void si_llvm_load_input_vs( tmp = ac_build_opencoded_load_format( &ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1, fix_fetch.u.format, fix_fetch.u.reverse, !opencode, - t_list, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, 0, true); + vb_desc, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, 0, true); for (unsigned i = 0; i < 4; ++i) out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), ""); return; @@ -526,7 +514,7 @@ void si_llvm_load_input_vs( for (unsigned i = 0; i < num_fetches; ++i) { LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0); - fetches[i] = ac_build_buffer_load_format(&ctx->ac, t_list, vertex_index, voffset, + fetches[i] = ac_build_buffer_load_format(&ctx->ac, vb_desc, vertex_index, voffset, channels_per_fetch, 0, true); } @@ -596,15 +584,6 @@ void si_llvm_load_input_vs( out[i] = ac_to_float(&ctx->ac, fetches[i]); } -static void declare_input_vs( - struct si_shader_context *ctx, - unsigned input_index, - const struct tgsi_full_declaration *decl, - LLVMValueRef out[4]) -{ - si_llvm_load_input_vs(ctx, input_index, out); -} - LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx, unsigned swizzle) { @@ -626,53 +605,6 @@ LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx, } } -/** - * Return the value of tgsi_ind_register for indexing. - * This is the indirect index with the constant offset added to it. - */ -LLVMValueRef si_get_indirect_index(struct si_shader_context *ctx, - const struct tgsi_ind_register *ind, - unsigned addr_mul, - int rel_index) -{ - LLVMValueRef result; - - if (ind->File == TGSI_FILE_ADDRESS) { - result = ctx->addrs[ind->Index][ind->Swizzle]; - result = LLVMBuildLoad(ctx->ac.builder, result, ""); - } else { - struct tgsi_full_src_register src = {}; - - src.Register.File = ind->File; - src.Register.Index = ind->Index; - - /* Set the second index to 0 for constants. */ - if (ind->File == TGSI_FILE_CONSTANT) - src.Register.Dimension = 1; - - result = ctx->bld_base.emit_fetch_funcs[ind->File](&ctx->bld_base, &src, - TGSI_TYPE_SIGNED, - ind->Swizzle); - result = ac_to_integer(&ctx->ac, result); - } - - return ac_build_imad(&ctx->ac, result, LLVMConstInt(ctx->i32, addr_mul, 0), - LLVMConstInt(ctx->i32, rel_index, 0)); -} - -/** - * Like si_get_indirect_index, but restricts the return value to a (possibly - * undefined) value inside [0..num). - */ -LLVMValueRef si_get_bounded_indirect_index(struct si_shader_context *ctx, - const struct tgsi_ind_register *ind, - int rel_index, unsigned num) -{ - LLVMValueRef result = si_get_indirect_index(ctx, ind, 1, rel_index); - - return si_llvm_bound_index(ctx, result, num); -} - static LLVMValueRef get_dw_address_from_generic_indices(struct si_shader_context *ctx, LLVMValueRef vertex_dw_stride, LLVMValueRef base_addr, @@ -701,78 +633,6 @@ static LLVMValueRef get_dw_address_from_generic_indices(struct si_shader_context LLVMConstInt(ctx->i32, param * 4, 0), ""); } -/** - * Calculate a dword address given an input or output register and a stride. - */ -static LLVMValueRef get_dw_address(struct si_shader_context *ctx, - const struct tgsi_full_dst_register *dst, - const struct tgsi_full_src_register *src, - LLVMValueRef vertex_dw_stride, - LLVMValueRef base_addr) -{ - struct tgsi_shader_info *info = &ctx->shader->selector->info; - ubyte *name, *index, *array_first; - int input_index; - struct tgsi_full_dst_register reg; - LLVMValueRef vertex_index = NULL; - LLVMValueRef ind_index = NULL; - - /* Set the register description. The address computation is the same - * for sources and destinations. */ - if (src) { - reg.Register.File = src->Register.File; - reg.Register.Index = src->Register.Index; - reg.Register.Indirect = src->Register.Indirect; - reg.Register.Dimension = src->Register.Dimension; - reg.Indirect = src->Indirect; - reg.Dimension = src->Dimension; - reg.DimIndirect = src->DimIndirect; - } else - reg = *dst; - - /* If the register is 2-dimensional (e.g. an array of vertices - * in a primitive), calculate the base address of the vertex. */ - if (reg.Register.Dimension) { - if (reg.Dimension.Indirect) - vertex_index = si_get_indirect_index(ctx, ®.DimIndirect, - 1, reg.Dimension.Index); - else - vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0); - } - - /* Get information about the register. */ - if (reg.Register.File == TGSI_FILE_INPUT) { - name = info->input_semantic_name; - index = info->input_semantic_index; - array_first = info->input_array_first; - } else if (reg.Register.File == TGSI_FILE_OUTPUT) { - name = info->output_semantic_name; - index = info->output_semantic_index; - array_first = info->output_array_first; - } else { - assert(0); - return NULL; - } - - if (reg.Register.Indirect) { - /* Add the relative address of the element. */ - if (reg.Indirect.ArrayID) - input_index = array_first[reg.Indirect.ArrayID]; - else - input_index = reg.Register.Index; - - ind_index = si_get_indirect_index(ctx, ®.Indirect, - 1, reg.Register.Index - input_index); - } else { - input_index = reg.Register.Index; - } - - return get_dw_address_from_generic_indices(ctx, vertex_dw_stride, - base_addr, vertex_index, - ind_index, name[input_index], - index[input_index]); -} - /* The offchip buffer layout for TCS->TES is * * - attribute 0 of patch 0 vertex 0 @@ -827,21 +687,19 @@ static LLVMValueRef get_tcs_tes_buffer_address(struct si_shader_context *ctx, return base_addr; } -/* This is a generic helper that can be shared by the NIR and TGSI backends */ static LLVMValueRef get_tcs_tes_buffer_address_from_generic_indices( struct si_shader_context *ctx, LLVMValueRef vertex_index, LLVMValueRef param_index, - unsigned param_base, - ubyte *name, - ubyte *index, - bool is_patch) + ubyte name, ubyte index) { unsigned param_index_base; - param_index_base = is_patch ? - si_shader_io_get_unique_index_patch(name[param_base], index[param_base]) : - si_shader_io_get_unique_index(name[param_base], index[param_base], false); + param_index_base = name == TGSI_SEMANTIC_PATCH || + name == TGSI_SEMANTIC_TESSINNER || + name == TGSI_SEMANTIC_TESSOUTER ? + si_shader_io_get_unique_index_patch(name, index) : + si_shader_io_get_unique_index(name, index, false); if (param_index) { param_index = LLVMBuildAdd(ctx->ac.builder, param_index, @@ -855,67 +713,24 @@ static LLVMValueRef get_tcs_tes_buffer_address_from_generic_indices( vertex_index, param_index); } -static LLVMValueRef get_tcs_tes_buffer_address_from_reg( - struct si_shader_context *ctx, - const struct tgsi_full_dst_register *dst, - const struct tgsi_full_src_register *src) +static LLVMValueRef si_build_gather_64bit(struct si_shader_context *ctx, + LLVMTypeRef type, + LLVMValueRef val1, + LLVMValueRef val2) { - struct tgsi_shader_info *info = &ctx->shader->selector->info; - ubyte *name, *index, *array_first; - struct tgsi_full_src_register reg; - LLVMValueRef vertex_index = NULL; - LLVMValueRef param_index = NULL; - unsigned param_base; - - reg = src ? *src : tgsi_full_src_register_from_dst(dst); - - if (reg.Register.Dimension) { - - if (reg.Dimension.Indirect) - vertex_index = si_get_indirect_index(ctx, ®.DimIndirect, - 1, reg.Dimension.Index); - else - vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0); - } - - /* Get information about the register. */ - if (reg.Register.File == TGSI_FILE_INPUT) { - name = info->input_semantic_name; - index = info->input_semantic_index; - array_first = info->input_array_first; - } else if (reg.Register.File == TGSI_FILE_OUTPUT) { - name = info->output_semantic_name; - index = info->output_semantic_index; - array_first = info->output_array_first; - } else { - assert(0); - return NULL; - } - - if (reg.Register.Indirect) { - if (reg.Indirect.ArrayID) - param_base = array_first[reg.Indirect.ArrayID]; - else - param_base = reg.Register.Index; - - param_index = si_get_indirect_index(ctx, ®.Indirect, - 1, reg.Register.Index - param_base); - - } else { - param_base = reg.Register.Index; - } - - return get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index, - param_index, param_base, - name, index, !reg.Register.Dimension); + LLVMValueRef values[2] = { + ac_to_integer(&ctx->ac, val1), + ac_to_integer(&ctx->ac, val2), + }; + LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, 2); + return LLVMBuildBitCast(ctx->ac.builder, result, type, ""); } -static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base, +static LLVMValueRef buffer_load(struct si_shader_context *ctx, LLVMTypeRef type, unsigned swizzle, LLVMValueRef buffer, LLVMValueRef offset, LLVMValueRef base, bool can_speculate) { - struct si_shader_context *ctx = si_shader_context(bld_base); LLVMValueRef value, value2; LLVMTypeRef vec_type = LLVMVectorType(type, 4); @@ -941,7 +756,7 @@ static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base, value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset, swizzle * 4 + 4, ac_glc, can_speculate, false); - return si_llvm_emit_fetch_64bit(bld_base, type, value, value2); + return si_build_gather_64bit(ctx, type, value, value2); } /** @@ -951,30 +766,28 @@ static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base, * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4 * \param dw_addr address in dwords */ -static LLVMValueRef lshs_lds_load(struct lp_build_tgsi_context *bld_base, - LLVMTypeRef type, unsigned swizzle, - LLVMValueRef dw_addr) +static LLVMValueRef lshs_lds_load(struct si_shader_context *ctx, + LLVMTypeRef type, unsigned swizzle, + LLVMValueRef dw_addr) { - struct si_shader_context *ctx = si_shader_context(bld_base); LLVMValueRef value; if (swizzle == ~0) { - LLVMValueRef values[TGSI_NUM_CHANNELS]; + LLVMValueRef values[4]; - for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++) - values[chan] = lshs_lds_load(bld_base, type, chan, dw_addr); + for (unsigned chan = 0; chan < 4; chan++) + values[chan] = lshs_lds_load(ctx, type, chan, dw_addr); - return ac_build_gather_values(&ctx->ac, values, - TGSI_NUM_CHANNELS); + return ac_build_gather_values(&ctx->ac, values, 4); } /* Split 64-bit loads. */ if (llvm_type_is_64bit(ctx, type)) { LLVMValueRef lo, hi; - lo = lshs_lds_load(bld_base, ctx->i32, swizzle, dw_addr); - hi = lshs_lds_load(bld_base, ctx->i32, swizzle + 1, dw_addr); - return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi); + lo = lshs_lds_load(ctx, ctx->i32, swizzle, dw_addr); + hi = lshs_lds_load(ctx, ctx->i32, swizzle + 1, dw_addr); + return si_build_gather_64bit(ctx, type, lo, hi); } dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr, @@ -1036,7 +849,7 @@ static LLVMValueRef get_tess_ring_descriptor(struct si_shader_context *ctx, if (ctx->screen->info.chip_class >= GFX10) rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) | - S_008F0C_OOB_SELECT(3) | + S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1); else rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | @@ -1052,21 +865,6 @@ static LLVMValueRef get_tess_ring_descriptor(struct si_shader_context *ctx, return ac_build_gather_values(&ctx->ac, desc, 4); } -static LLVMValueRef fetch_input_tcs( - struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_src_register *reg, - enum tgsi_opcode_type type, unsigned swizzle_in) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMValueRef dw_addr, stride; - unsigned swizzle = swizzle_in & 0xffff; - stride = get_tcs_in_vertex_dw_stride(ctx); - dw_addr = get_tcs_in_current_patch_offset(ctx); - dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr); - - return lshs_lds_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle, dw_addr); -} - static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi, LLVMTypeRef type, LLVMValueRef vertex_index, @@ -1081,12 +879,24 @@ static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi, bool load_input) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info *info = &ctx->shader->selector->info; - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; + struct si_shader_info *info = &ctx->shader->selector->info; LLVMValueRef dw_addr, stride; + ubyte name, index; driver_location = driver_location / 4; + if (load_input) { + name = info->input_semantic_name[driver_location]; + index = info->input_semantic_index[driver_location]; + } else { + name = info->output_semantic_name[driver_location]; + index = info->output_semantic_index[driver_location]; + } + + assert((name == TGSI_SEMANTIC_PATCH || + name == TGSI_SEMANTIC_TESSINNER || + name == TGSI_SEMANTIC_TESSOUTER) == is_patch); + if (load_input) { stride = get_tcs_in_vertex_dw_stride(ctx); dw_addr = get_tcs_in_current_patch_offset(ctx); @@ -1104,16 +914,6 @@ static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi, param_index = LLVMConstInt(ctx->i32, const_index, 0); } - ubyte name; - ubyte index; - if (load_input) { - name = info->input_semantic_name[driver_location]; - index = info->input_semantic_index[driver_location]; - } else { - name = info->output_semantic_name[driver_location]; - index = info->output_semantic_index[driver_location]; - } - dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr, vertex_index, param_index, name, index); @@ -1125,49 +925,12 @@ static LLVMValueRef si_nir_load_tcs_varyings(struct ac_shader_abi *abi, offset *= 2; offset += component; - value[i + component] = lshs_lds_load(bld_base, type, offset, dw_addr); + value[i + component] = lshs_lds_load(ctx, type, offset, dw_addr); } return ac_build_varying_gather_values(&ctx->ac, value, num_components, component); } -static LLVMValueRef fetch_output_tcs( - struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_src_register *reg, - enum tgsi_opcode_type type, unsigned swizzle_in) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMValueRef dw_addr, stride; - unsigned swizzle = (swizzle_in & 0xffff); - - if (reg->Register.Dimension) { - stride = get_tcs_out_vertex_dw_stride(ctx); - dw_addr = get_tcs_out_current_patch_offset(ctx); - dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr); - } else { - dw_addr = get_tcs_out_current_patch_data_offset(ctx); - dw_addr = get_dw_address(ctx, NULL, reg, NULL, dw_addr); - } - - return lshs_lds_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle, dw_addr); -} - -static LLVMValueRef fetch_input_tes( - struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_src_register *reg, - enum tgsi_opcode_type type, unsigned swizzle_in) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMValueRef base, addr; - unsigned swizzle = (swizzle_in & 0xffff); - - base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset); - addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg); - - return buffer_load(bld_base, tgsi2llvmtype(bld_base, type), swizzle, - ctx->tess_offchip_ring, base, addr, true); -} - LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi, LLVMTypeRef type, LLVMValueRef vertex_index, @@ -1182,10 +945,16 @@ LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi, bool load_input) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info *info = &ctx->shader->selector->info; + struct si_shader_info *info = &ctx->shader->selector->info; LLVMValueRef base, addr; driver_location = driver_location / 4; + ubyte name = info->input_semantic_name[driver_location]; + ubyte index = info->input_semantic_index[driver_location]; + + assert((name == TGSI_SEMANTIC_PATCH || + name == TGSI_SEMANTIC_TESSINNER || + name == TGSI_SEMANTIC_TESSOUTER) == is_patch); base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset); @@ -1194,15 +963,12 @@ LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi, } addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index, - param_index, driver_location, - info->input_semantic_name, - info->input_semantic_index, - is_patch); + param_index, + name, index); /* TODO: This will generate rather ordinary llvm code, although it * should be easy for the optimiser to fix up. In future we might want - * to refactor buffer_load(), but for now this maximises code sharing - * between the NIR and TGSI backends. + * to refactor buffer_load(). */ LLVMValueRef value[4]; for (unsigned i = 0; i < num_components; i++) { @@ -1210,123 +976,25 @@ LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi, if (llvm_type_is_64bit(ctx, type)) { offset *= 2; if (offset == 4) { + ubyte name = info->input_semantic_name[driver_location + 1]; + ubyte index = info->input_semantic_index[driver_location + 1]; addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index, param_index, - driver_location + 1, - info->input_semantic_name, - info->input_semantic_index, - is_patch); + name, index); } offset = offset % 4; } offset += component; - value[i + component] = buffer_load(&ctx->bld_base, type, offset, + value[i + component] = buffer_load(ctx, type, offset, ctx->tess_offchip_ring, base, addr, true); } return ac_build_varying_gather_values(&ctx->ac, value, num_components, component); } -static void store_output_tcs(struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_instruction *inst, - const struct tgsi_opcode_info *info, - unsigned index, - LLVMValueRef dst[4]) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - const struct tgsi_full_dst_register *reg = &inst->Dst[index]; - const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info; - unsigned chan_index; - LLVMValueRef dw_addr, stride; - LLVMValueRef buffer, base, buf_addr; - LLVMValueRef values[4]; - bool skip_lds_store; - bool is_tess_factor = false, is_tess_inner = false; - - /* Only handle per-patch and per-vertex outputs here. - * Vectors will be lowered to scalars and this function will be called again. - */ - if (reg->Register.File != TGSI_FILE_OUTPUT || - (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) { - si_llvm_emit_store(bld_base, inst, info, index, dst); - return; - } - - if (reg->Register.Dimension) { - stride = get_tcs_out_vertex_dw_stride(ctx); - dw_addr = get_tcs_out_current_patch_offset(ctx); - dw_addr = get_dw_address(ctx, reg, NULL, stride, dw_addr); - skip_lds_store = !sh_info->reads_pervertex_outputs; - } else { - dw_addr = get_tcs_out_current_patch_data_offset(ctx); - dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr); - skip_lds_store = !sh_info->reads_perpatch_outputs; - - if (!reg->Register.Indirect) { - int name = sh_info->output_semantic_name[reg->Register.Index]; - - /* Always write tess factors into LDS for the TCS epilog. */ - if (name == TGSI_SEMANTIC_TESSINNER || - name == TGSI_SEMANTIC_TESSOUTER) { - /* The epilog doesn't read LDS if invocation 0 defines tess factors. */ - skip_lds_store = !sh_info->reads_tessfactor_outputs && - ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs; - is_tess_factor = true; - is_tess_inner = name == TGSI_SEMANTIC_TESSINNER; - } - } - } - - buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS); - - base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset); - buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL); - - uint32_t writemask = reg->Register.WriteMask; - while (writemask) { - chan_index = u_bit_scan(&writemask); - LLVMValueRef value = dst[chan_index]; - - if (inst->Instruction.Saturate) - value = ac_build_clamp(&ctx->ac, value); - - /* Skip LDS stores if there is no LDS read of this output. */ - if (!skip_lds_store) - lshs_lds_store(ctx, chan_index, dw_addr, value); - - value = ac_to_integer(&ctx->ac, value); - values[chan_index] = value; - - if (reg->Register.WriteMask != 0xF && !is_tess_factor) { - ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1, - buf_addr, base, - 4 * chan_index, ac_glc); - } - - /* Write tess factors into VGPRs for the epilog. */ - if (is_tess_factor && - ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) { - if (!is_tess_inner) { - LLVMBuildStore(ctx->ac.builder, value, /* outer */ - ctx->invoc0_tess_factors[chan_index]); - } else if (chan_index < 2) { - LLVMBuildStore(ctx->ac.builder, value, /* inner */ - ctx->invoc0_tess_factors[4 + chan_index]); - } - } - } - - if (reg->Register.WriteMask == 0xF && !is_tess_factor) { - LLVMValueRef value = ac_build_gather_values(&ctx->ac, - values, 4); - ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buf_addr, - base, 0, ac_glc); - } -} - static void si_nir_store_output_tcs(struct ac_shader_abi *abi, const struct nir_variable *var, LLVMValueRef vertex_index, @@ -1336,9 +1004,8 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, unsigned writemask) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info *info = &ctx->shader->selector->info; + struct si_shader_info *info = &ctx->shader->selector->info; const unsigned component = var->data.location_frac; - const bool is_patch = var->data.patch; unsigned driver_location = var->data.driver_location; LLVMValueRef dw_addr, stride; LLVMValueRef buffer, base, addr; @@ -1354,6 +1021,14 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, if (!param_index) param_index = LLVMConstInt(ctx->i32, const_index, 0); + const bool is_patch = var->data.patch || + var->data.location == VARYING_SLOT_TESS_LEVEL_INNER || + var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER; + + assert((name == TGSI_SEMANTIC_PATCH || + name == TGSI_SEMANTIC_TESSINNER || + name == TGSI_SEMANTIC_TESSOUTER) == is_patch); + if (!is_patch) { stride = get_tcs_out_vertex_dw_stride(ctx); dw_addr = get_tcs_out_current_patch_offset(ctx); @@ -1378,7 +1053,7 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, name == TGSI_SEMANTIC_TESSOUTER) { /* The epilog doesn't read LDS if invocation 0 defines tess factors. */ skip_lds_store = !info->reads_tessfactor_outputs && - ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs; + ctx->shader->selector->info.tessfactors_are_def_in_all_invocs; is_tess_factor = true; is_tess_inner = name == TGSI_SEMANTIC_TESSINNER; } @@ -1390,25 +1065,21 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset); addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index, - param_index, driver_location, - info->output_semantic_name, - info->output_semantic_index, - is_patch); + param_index, name, index); - for (unsigned chan = 0; chan < 8; chan++) { + for (unsigned chan = component; chan < 8; chan++) { if (!(writemask & (1 << chan))) continue; LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component); unsigned buffer_store_offset = chan % 4; if (chan == 4) { + ubyte name = info->output_semantic_name[driver_location + 1]; + ubyte index = info->output_semantic_index[driver_location + 1]; addr = get_tcs_tes_buffer_address_from_generic_indices(ctx, vertex_index, param_index, - driver_location + 1, - info->output_semantic_name, - info->output_semantic_index, - is_patch); + name, index); } /* Skip LDS stores if there is no LDS read of this output. */ @@ -1427,7 +1098,7 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, /* Write tess factors into VGPRs for the epilog. */ if (is_tess_factor && - ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) { + ctx->shader->selector->info.tessfactors_are_def_in_all_invocs) { if (!is_tess_inner) { LLVMBuildStore(ctx->ac.builder, value, /* outer */ ctx->invoc0_tess_factors[chan]); @@ -1446,17 +1117,16 @@ static void si_nir_store_output_tcs(struct ac_shader_abi *abi, } } -LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi, - unsigned input_index, - unsigned vtx_offset_param, - LLVMTypeRef type, - unsigned swizzle) +static LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi, + unsigned input_index, + unsigned vtx_offset_param, + LLVMTypeRef type, + unsigned swizzle) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; struct si_shader *shader = ctx->shader; LLVMValueRef vtx_offset, soffset; - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; unsigned semantic_name = info->input_semantic_name[input_index]; unsigned semantic_index = info->input_semantic_index[input_index]; unsigned param; @@ -1506,14 +1176,13 @@ LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi, /* GFX6: input load from the ESGS ring in memory. */ if (swizzle == ~0) { - LLVMValueRef values[TGSI_NUM_CHANNELS]; + LLVMValueRef values[4]; unsigned chan; - for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { + for (chan = 0; chan < 4; chan++) { values[chan] = si_llvm_load_input_gs(abi, input_index, vtx_offset_param, type, chan); } - return ac_build_gather_values(&ctx->ac, values, - TGSI_NUM_CHANNELS); + return ac_build_gather_values(&ctx->ac, values, 4); } /* Get the vertex offset parameter on GFX6. */ @@ -1534,7 +1203,7 @@ LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi, value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0, vtx_offset, soffset, 0, ac_glc, true, false); - return si_llvm_emit_fetch_64bit(bld_base, type, value, value2); + return si_build_gather_64bit(ctx, type, value, value2); } return LLVMBuildBitCast(ctx->ac.builder, value, type, ""); } @@ -1564,58 +1233,6 @@ static LLVMValueRef si_nir_load_input_gs(struct ac_shader_abi *abi, return ac_build_varying_gather_values(&ctx->ac, value, num_components, component); } -static LLVMValueRef fetch_input_gs( - struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_src_register *reg, - enum tgsi_opcode_type type, - unsigned swizzle_in) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - struct tgsi_shader_info *info = &ctx->shader->selector->info; - unsigned swizzle = swizzle_in & 0xffff; - - unsigned semantic_name = info->input_semantic_name[reg->Register.Index]; - if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID) - return si_get_primitive_id(ctx, swizzle); - - if (!reg->Register.Dimension) - return NULL; - - return si_llvm_load_input_gs(&ctx->abi, reg->Register.Index, - reg->Dimension.Index, - tgsi2llvmtype(bld_base, type), - swizzle); -} - -static int lookup_interp_param_index(unsigned interpolate, unsigned location) -{ - switch (interpolate) { - case TGSI_INTERPOLATE_CONSTANT: - return 0; - - case TGSI_INTERPOLATE_LINEAR: - if (location == TGSI_INTERPOLATE_LOC_SAMPLE) - return SI_PARAM_LINEAR_SAMPLE; - else if (location == TGSI_INTERPOLATE_LOC_CENTROID) - return SI_PARAM_LINEAR_CENTROID; - else - return SI_PARAM_LINEAR_CENTER; - break; - case TGSI_INTERPOLATE_COLOR: - case TGSI_INTERPOLATE_PERSPECTIVE: - if (location == TGSI_INTERPOLATE_LOC_SAMPLE) - return SI_PARAM_PERSP_SAMPLE; - else if (location == TGSI_INTERPOLATE_LOC_CENTROID) - return SI_PARAM_PERSP_CENTROID; - else - return SI_PARAM_PERSP_CENTER; - break; - default: - fprintf(stderr, "Warning: Unhandled interpolation mode.\n"); - return -1; - } -} - static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx, unsigned attr_index, unsigned chan, LLVMValueRef prim_mask, @@ -1648,9 +1265,8 @@ static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx, * @param face SI_PARAM_FRONT_FACE * @param result the return value (4 components) */ -static void interp_fs_input(struct si_shader_context *ctx, +static void interp_fs_color(struct si_shader_context *ctx, unsigned input_index, - unsigned semantic_name, unsigned semantic_index, unsigned num_interp_inputs, unsigned colors_read_mask, @@ -1687,8 +1303,7 @@ static void interp_fs_input(struct si_shader_context *ctx, ctx->i32_1, ""); } - if (semantic_name == TGSI_SEMANTIC_COLOR && - ctx->shader->key.part.ps.prolog.color_two_side) { + if (ctx->shader->key.part.ps.prolog.color_two_side) { LLVMValueRef is_face_positive; /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1", @@ -1701,7 +1316,7 @@ static void interp_fs_input(struct si_shader_context *ctx, is_face_positive = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, face, ctx->i32_0, ""); - for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { + for (chan = 0; chan < 4; chan++) { LLVMValueRef front, back; front = si_build_fs_interp(ctx, @@ -1717,14 +1332,8 @@ static void interp_fs_input(struct si_shader_context *ctx, back, ""); } - } else if (semantic_name == TGSI_SEMANTIC_FOG) { - result[0] = si_build_fs_interp(ctx, input_index, - 0, prim_mask, i, j); - result[1] = - result[2] = LLVMConstReal(ctx->f32, 0.0f); - result[3] = LLVMConstReal(ctx->f32, 1.0f); } else { - for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { + for (chan = 0; chan < 4; chan++) { result[chan] = si_build_fs_interp(ctx, input_index, chan, prim_mask, i, j); @@ -1732,60 +1341,6 @@ static void interp_fs_input(struct si_shader_context *ctx, } } -void si_llvm_load_input_fs( - struct si_shader_context *ctx, - unsigned input_index, - LLVMValueRef out[4]) -{ - struct si_shader *shader = ctx->shader; - struct tgsi_shader_info *info = &shader->selector->info; - LLVMValueRef main_fn = ctx->main_fn; - LLVMValueRef interp_param = NULL; - int interp_param_idx; - enum tgsi_semantic semantic_name = info->input_semantic_name[input_index]; - unsigned semantic_index = info->input_semantic_index[input_index]; - enum tgsi_interpolate_mode interp_mode = info->input_interpolate[input_index]; - enum tgsi_interpolate_loc interp_loc = info->input_interpolate_loc[input_index]; - - /* Get colors from input VGPRs (set by the prolog). */ - if (semantic_name == TGSI_SEMANTIC_COLOR) { - unsigned colors_read = shader->selector->info.colors_read; - unsigned mask = colors_read >> (semantic_index * 4); - unsigned offset = SI_PARAM_POS_FIXED_PT + 1 + - (semantic_index ? util_bitcount(colors_read & 0xf) : 0); - LLVMValueRef undef = LLVMGetUndef(ctx->f32); - - out[0] = mask & 0x1 ? LLVMGetParam(main_fn, offset++) : undef; - out[1] = mask & 0x2 ? LLVMGetParam(main_fn, offset++) : undef; - out[2] = mask & 0x4 ? LLVMGetParam(main_fn, offset++) : undef; - out[3] = mask & 0x8 ? LLVMGetParam(main_fn, offset++) : undef; - return; - } - - interp_param_idx = lookup_interp_param_index(interp_mode, interp_loc); - if (interp_param_idx == -1) - return; - else if (interp_param_idx) { - interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx); - } - - interp_fs_input(ctx, input_index, semantic_name, - semantic_index, 0, /* this param is unused */ - shader->selector->info.colors_read, interp_param, - ac_get_arg(&ctx->ac, ctx->args.prim_mask), - LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE), - &out[0]); -} - -static void declare_input_fs( - struct si_shader_context *ctx, - unsigned input_index, - const struct tgsi_full_declaration *decl, - LLVMValueRef out[4]) -{ - si_llvm_load_input_fs(ctx, input_index, out); -} - LLVMValueRef si_get_sample_id(struct si_shader_context *ctx) { return si_unpack_param(ctx, ctx->args.ancillary, 8, 4); @@ -1907,7 +1462,7 @@ static LLVMValueRef load_tess_level(struct si_shader_context *ctx, addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL, LLVMConstInt(ctx->i32, param, 0)); - return buffer_load(&ctx->bld_base, ctx->f32, + return buffer_load(ctx, ctx->f32, ~0, ctx->tess_offchip_ring, base, addr, true); } @@ -1938,247 +1493,42 @@ static LLVMValueRef si_load_tess_level(struct ac_shader_abi *abi, if (load_default_state) { switch (varying_id) { - case VARYING_SLOT_TESS_LEVEL_INNER: - semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL; - break; - case VARYING_SLOT_TESS_LEVEL_OUTER: - semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL; - break; - default: - unreachable("unknown tess level"); - } - return load_tess_level_default(ctx, semantic_name); - } - - switch (varying_id) { - case VARYING_SLOT_TESS_LEVEL_INNER: - semantic_name = TGSI_SEMANTIC_TESSINNER; - break; - case VARYING_SLOT_TESS_LEVEL_OUTER: - semantic_name = TGSI_SEMANTIC_TESSOUTER; - break; - default: - unreachable("unknown tess level"); - } - - return load_tess_level(ctx, semantic_name); - -} - -static LLVMValueRef si_load_patch_vertices_in(struct ac_shader_abi *abi) -{ - struct si_shader_context *ctx = si_shader_context_from_abi(abi); - if (ctx->type == PIPE_SHADER_TESS_CTRL) - return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 13, 6); - else if (ctx->type == PIPE_SHADER_TESS_EVAL) - return get_num_tcs_out_vertices(ctx); - else - unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN"); -} - -void si_load_system_value(struct si_shader_context *ctx, - unsigned index, - const struct tgsi_full_declaration *decl) -{ - LLVMValueRef value = 0; - - assert(index < RADEON_LLVM_MAX_SYSTEM_VALUES); - - switch (decl->Semantic.Name) { - case TGSI_SEMANTIC_INSTANCEID: - value = ctx->abi.instance_id; - break; - - case TGSI_SEMANTIC_VERTEXID: - value = LLVMBuildAdd(ctx->ac.builder, - ctx->abi.vertex_id, - ac_get_arg(&ctx->ac, ctx->args.base_vertex), ""); - break; - - case TGSI_SEMANTIC_VERTEXID_NOBASE: - /* Unused. Clarify the meaning in indexed vs. non-indexed - * draws if this is ever used again. */ - assert(false); - break; - - case TGSI_SEMANTIC_BASEVERTEX: - value = get_base_vertex(&ctx->abi); - break; - - case TGSI_SEMANTIC_BASEINSTANCE: - value = ac_get_arg(&ctx->ac, ctx->args.start_instance); - break; - - case TGSI_SEMANTIC_DRAWID: - value = ac_get_arg(&ctx->ac, ctx->args.draw_id); - break; - - case TGSI_SEMANTIC_INVOCATIONID: - if (ctx->type == PIPE_SHADER_TESS_CTRL) { - value = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5); - } else if (ctx->type == PIPE_SHADER_GEOMETRY) { - if (ctx->screen->info.chip_class >= GFX10) { - value = LLVMBuildAnd(ctx->ac.builder, - ac_get_arg(&ctx->ac, ctx->args.gs_invocation_id), - LLVMConstInt(ctx->i32, 127, 0), ""); - } else { - value = ac_get_arg(&ctx->ac, ctx->args.gs_invocation_id); - } - } else { - assert(!"INVOCATIONID not implemented"); - } - break; - - case TGSI_SEMANTIC_POSITION: - { - LLVMValueRef pos[4] = { - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT), - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT), - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT), - ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_W_FLOAT)), - }; - value = ac_build_gather_values(&ctx->ac, pos, 4); - break; - } - - case TGSI_SEMANTIC_FACE: - value = ac_get_arg(&ctx->ac, ctx->args.front_face); - break; - - case TGSI_SEMANTIC_SAMPLEID: - value = si_get_sample_id(ctx); - break; - - case TGSI_SEMANTIC_SAMPLEPOS: { - LLVMValueRef pos[4] = { - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT), - LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT), - LLVMConstReal(ctx->f32, 0), - LLVMConstReal(ctx->f32, 0) - }; - pos[0] = ac_build_fract(&ctx->ac, pos[0], 32); - pos[1] = ac_build_fract(&ctx->ac, pos[1], 32); - value = ac_build_gather_values(&ctx->ac, pos, 4); - break; - } - - case TGSI_SEMANTIC_SAMPLEMASK: - /* This can only occur with the OpenGL Core profile, which - * doesn't support smoothing. - */ - value = LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE); - break; - - case TGSI_SEMANTIC_TESSCOORD: - value = si_load_tess_coord(&ctx->abi); - break; - - case TGSI_SEMANTIC_VERTICESIN: - value = si_load_patch_vertices_in(&ctx->abi); - break; - - case TGSI_SEMANTIC_TESSINNER: - case TGSI_SEMANTIC_TESSOUTER: - value = load_tess_level(ctx, decl->Semantic.Name); - break; - - case TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL: - case TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL: - value = load_tess_level_default(ctx, decl->Semantic.Name); - break; - - case TGSI_SEMANTIC_PRIMID: - value = si_get_primitive_id(ctx, 0); - break; - - case TGSI_SEMANTIC_GRID_SIZE: - value = ac_get_arg(&ctx->ac, ctx->args.num_work_groups); - break; - - case TGSI_SEMANTIC_BLOCK_SIZE: - value = get_block_size(&ctx->abi); - break; - - case TGSI_SEMANTIC_BLOCK_ID: - { - LLVMValueRef values[3]; - - for (int i = 0; i < 3; i++) { - values[i] = ctx->i32_0; - if (ctx->args.workgroup_ids[i].used) { - values[i] = ac_get_arg(&ctx->ac, ctx->args.workgroup_ids[i]); - } - } - value = ac_build_gather_values(&ctx->ac, values, 3); - break; - } - - case TGSI_SEMANTIC_THREAD_ID: - value = ac_get_arg(&ctx->ac, ctx->args.local_invocation_ids); - break; - - case TGSI_SEMANTIC_HELPER_INVOCATION: - value = ac_build_load_helper_invocation(&ctx->ac); - break; - - case TGSI_SEMANTIC_SUBGROUP_SIZE: - value = LLVMConstInt(ctx->i32, ctx->ac.wave_size, 0); - break; - - case TGSI_SEMANTIC_SUBGROUP_INVOCATION: - value = ac_get_thread_id(&ctx->ac); - break; - - case TGSI_SEMANTIC_SUBGROUP_EQ_MASK: - { - LLVMValueRef id = ac_get_thread_id(&ctx->ac); - if (ctx->ac.wave_size == 64) - id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, ""); - value = LLVMBuildShl(ctx->ac.builder, - LLVMConstInt(ctx->ac.iN_wavemask, 1, 0), id, ""); - if (ctx->ac.wave_size == 32) - value = LLVMBuildZExt(ctx->ac.builder, value, ctx->i64, ""); - value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->v2i32, ""); - break; - } - - case TGSI_SEMANTIC_SUBGROUP_GE_MASK: - case TGSI_SEMANTIC_SUBGROUP_GT_MASK: - case TGSI_SEMANTIC_SUBGROUP_LE_MASK: - case TGSI_SEMANTIC_SUBGROUP_LT_MASK: - { - LLVMValueRef id = ac_get_thread_id(&ctx->ac); - if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_GT_MASK || - decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK) { - /* All bits set except LSB */ - value = LLVMConstInt(ctx->ac.iN_wavemask, -2, 0); - } else { - /* All bits set */ - value = LLVMConstInt(ctx->ac.iN_wavemask, -1, 0); + case VARYING_SLOT_TESS_LEVEL_INNER: + semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL; + break; + case VARYING_SLOT_TESS_LEVEL_OUTER: + semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL; + break; + default: + unreachable("unknown tess level"); } - if (ctx->ac.wave_size == 64) - id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, ""); - value = LLVMBuildShl(ctx->ac.builder, value, id, ""); - if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK || - decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LT_MASK) - value = LLVMBuildNot(ctx->ac.builder, value, ""); - if (ctx->ac.wave_size == 32) - value = LLVMBuildZExt(ctx->ac.builder, value, ctx->i64, ""); - value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->v2i32, ""); - break; + return load_tess_level_default(ctx, semantic_name); } - case TGSI_SEMANTIC_CS_USER_DATA_AMD: - value = ac_get_arg(&ctx->ac, ctx->cs_user_data); + switch (varying_id) { + case VARYING_SLOT_TESS_LEVEL_INNER: + semantic_name = TGSI_SEMANTIC_TESSINNER; + break; + case VARYING_SLOT_TESS_LEVEL_OUTER: + semantic_name = TGSI_SEMANTIC_TESSOUTER; break; - default: - assert(!"unknown system value"); - return; + unreachable("unknown tess level"); } - ctx->system_values[index] = value; + return load_tess_level(ctx, semantic_name); + +} + +static LLVMValueRef si_load_patch_vertices_in(struct ac_shader_abi *abi) +{ + struct si_shader_context *ctx = si_shader_context_from_abi(abi); + if (ctx->type == PIPE_SHADER_TESS_CTRL) + return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 13, 6); + else if (ctx->type == PIPE_SHADER_TESS_EVAL) + return get_num_tcs_out_vertices(ctx); + else + unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN"); } void si_declare_compute_memory(struct si_shader_context *ctx) @@ -2200,15 +1550,6 @@ void si_declare_compute_memory(struct si_shader_context *ctx) ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, var, i8p, ""); } -void si_tgsi_declare_compute_memory(struct si_shader_context *ctx, - const struct tgsi_full_declaration *decl) -{ - assert(decl->Declaration.MemType == TGSI_MEMORY_TYPE_SHARED); - assert(decl->Range.First == decl->Range.Last); - - si_declare_compute_memory(ctx); -} - static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *ctx) { LLVMValueRef ptr = @@ -2234,7 +1575,7 @@ static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *c if (ctx->screen->info.chip_class >= GFX10) rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) | - S_008F0C_OOB_SELECT(3) | + S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1); else rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | @@ -2243,22 +1584,13 @@ static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *c LLVMValueRef desc_elems[] = { desc0, desc1, - LLVMConstInt(ctx->i32, (sel->info.const_file_max[0] + 1) * 16, 0), + LLVMConstInt(ctx->i32, sel->info.constbuf0_num_slots * 16, 0), LLVMConstInt(ctx->i32, rsrc3, false) }; return ac_build_gather_values(&ctx->ac, desc_elems, 4); } -static LLVMValueRef load_const_buffer_desc(struct si_shader_context *ctx, int i) -{ - LLVMValueRef list_ptr = ac_get_arg(&ctx->ac, - ctx->const_and_shader_buffers); - - return ac_build_load_to_sgpr(&ctx->ac, list_ptr, - LLVMConstInt(ctx->i32, si_get_constbuf_slot(i), 0)); -} - static LLVMValueRef load_ubo(struct ac_shader_abi *abi, LLVMValueRef index) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); @@ -2293,72 +1625,6 @@ load_ssbo(struct ac_shader_abi *abi, LLVMValueRef index, bool write) return ac_build_load_to_sgpr(&ctx->ac, rsrc_ptr, index); } -static LLVMValueRef fetch_constant( - struct lp_build_tgsi_context *bld_base, - const struct tgsi_full_src_register *reg, - enum tgsi_opcode_type type, - unsigned swizzle_in) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - struct si_shader_selector *sel = ctx->shader->selector; - const struct tgsi_ind_register *ireg = ®->Indirect; - unsigned buf, idx; - unsigned swizzle = swizzle_in & 0xffff; - - LLVMValueRef addr, bufp; - - if (swizzle_in == LP_CHAN_ALL) { - unsigned chan; - LLVMValueRef values[4]; - for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) - values[chan] = fetch_constant(bld_base, reg, type, chan); - - return ac_build_gather_values(&ctx->ac, values, 4); - } - - /* Split 64-bit loads. */ - if (tgsi_type_is_64bit(type)) { - LLVMValueRef lo, hi; - - lo = fetch_constant(bld_base, reg, TGSI_TYPE_UNSIGNED, swizzle); - hi = fetch_constant(bld_base, reg, TGSI_TYPE_UNSIGNED, (swizzle_in >> 16)); - return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type), - lo, hi); - } - - idx = reg->Register.Index * 4 + swizzle; - if (reg->Register.Indirect) { - addr = si_get_indirect_index(ctx, ireg, 16, idx * 4); - } else { - addr = LLVMConstInt(ctx->i32, idx * 4, 0); - } - - /* Fast path when user data SGPRs point to constant buffer 0 directly. */ - if (sel->info.const_buffers_declared == 1 && - sel->info.shader_buffers_declared == 0) { - LLVMValueRef desc = load_const_buffer_desc_fast_path(ctx); - LLVMValueRef result = buffer_load_const(ctx, desc, addr); - return bitcast(bld_base, type, result); - } - - assert(reg->Register.Dimension); - buf = reg->Dimension.Index; - - if (reg->Dimension.Indirect) { - LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers); - LLVMValueRef index; - index = si_get_bounded_indirect_index(ctx, ®->DimIndirect, - reg->Dimension.Index, - ctx->num_const_buffers); - index = LLVMBuildAdd(ctx->ac.builder, index, - LLVMConstInt(ctx->i32, SI_NUM_SHADER_BUFFERS, 0), ""); - bufp = ac_build_load_to_sgpr(&ctx->ac, ptr, index); - } else - bufp = load_const_buffer_desc(ctx, buf); - - return bitcast(bld_base, type, buffer_load_const(ctx, bufp, addr)); -} - /* Initialize arguments for the shader export intrinsic */ static void si_llvm_init_export_args(struct si_shader_context *ctx, LLVMValueRef *values, @@ -2489,11 +1755,8 @@ static void si_llvm_init_export_args(struct si_shader_context *ctx, } } -static void si_alpha_test(struct lp_build_tgsi_context *bld_base, - LLVMValueRef alpha) +static void si_alpha_test(struct si_shader_context *ctx, LLVMValueRef alpha) { - struct si_shader_context *ctx = si_shader_context(bld_base); - if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) { static LLVMRealPredicate cond_map[PIPE_FUNC_ALWAYS + 1] = { [PIPE_FUNC_LESS] = LLVMRealOLT, @@ -2516,11 +1779,10 @@ static void si_alpha_test(struct lp_build_tgsi_context *bld_base, } } -static LLVMValueRef si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context *bld_base, +static LLVMValueRef si_scale_alpha_by_sample_mask(struct si_shader_context *ctx, LLVMValueRef alpha, unsigned samplemask_param) { - struct si_shader_context *ctx = si_shader_context(bld_base); LLVMValueRef coverage; /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */ @@ -2563,8 +1825,8 @@ static void si_llvm_emit_clipvertex(struct si_shader_context *ctx, args->out[3] = LLVMConstReal(ctx->f32, 0.0f); /* Compute dot products of position and user clip plane vectors */ - for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { - for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) { + for (chan = 0; chan < 4; chan++) { + for (const_chan = 0; const_chan < 4; const_chan++) { LLVMValueRef addr = LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 + const_chan) * 4, 0); @@ -3024,9 +2286,8 @@ void si_llvm_export_vs(struct si_shader_context *ctx, * Forward all outputs from the vertex shader to the TES. This is only used * for the fixed function TCS. */ -static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base) +static void si_copy_tcs_inputs(struct si_shader_context *ctx) { - struct si_shader_context *ctx = si_shader_context(bld_base); LLVMValueRef invocation_id, buffer, buffer_offset; LLVMValueRef lds_vertex_stride, lds_base; uint64_t inputs; @@ -3053,21 +2314,20 @@ static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base) invocation_id, LLVMConstInt(ctx->i32, i, 0)); - LLVMValueRef value = lshs_lds_load(bld_base, ctx->ac.i32, ~0, lds_ptr); + LLVMValueRef value = lshs_lds_load(ctx, ctx->ac.i32, ~0, lds_ptr); ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr, buffer_offset, 0, ac_glc); } } -static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base, +static void si_write_tess_factors(struct si_shader_context *ctx, LLVMValueRef rel_patch_id, LLVMValueRef invocation_id, LLVMValueRef tcs_out_current_patch_data_offset, LLVMValueRef invoc0_tf_outer[4], LLVMValueRef invoc0_tf_inner[2]) { - struct si_shader_context *ctx = si_shader_context(bld_base); struct si_shader *shader = ctx->shader; unsigned tess_inner_index, tess_outer_index; LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer; @@ -3076,7 +2336,7 @@ static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base, /* Add a barrier before loading tess factors from LDS. */ if (!shader->key.part.tcs.epilog.invoc0_tess_factors_are_def) - si_llvm_emit_barrier(NULL, bld_base, NULL); + si_llvm_emit_barrier(ctx); /* Do this only for invocation 0, because the tess levels are per-patch, * not per-vertex. @@ -3138,17 +2398,17 @@ static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base, for (i = 0; i < outer_comps; i++) { outer[i] = out[i] = - lshs_lds_load(bld_base, ctx->ac.i32, i, lds_outer); + lshs_lds_load(ctx, ctx->ac.i32, i, lds_outer); } for (i = 0; i < inner_comps; i++) { inner[i] = out[outer_comps+i] = - lshs_lds_load(bld_base, ctx->ac.i32, i, lds_inner); + lshs_lds_load(ctx, ctx->ac.i32, i, lds_inner); } } if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) { /* For isolines, the hardware expects tess factors in the - * reverse order from what GLSL / TGSI specify. + * reverse order from what NIR specifies. */ LLVMValueRef tmp = out[0]; out[0] = out[1]; @@ -3273,11 +2533,10 @@ static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi *abi, LLVMValueRef *addrs) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; LLVMBuilderRef builder = ctx->ac.builder; LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset; - si_copy_tcs_inputs(bld_base); + si_copy_tcs_inputs(ctx); rel_patch_id = get_rel_patch_id(ctx); invocation_id = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5); @@ -3345,7 +2604,7 @@ static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi *abi, ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, ""); ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, ""); - if (ctx->shader->selector->tcs_info.tessfactors_are_def_in_all_invocs) { + if (ctx->shader->selector->info.tessfactors_are_def_in_all_invocs) { vgpr++; /* skip the tess factor LDS offset */ for (unsigned i = 0; i < 6; i++) { LLVMValueRef value = @@ -3443,7 +2702,7 @@ static void si_llvm_emit_ls_epilogue(struct ac_shader_abi *abi, { struct si_shader_context *ctx = si_shader_context_from_abi(abi); struct si_shader *shader = ctx->shader; - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; unsigned i, chan; LLVMValueRef vertex_id = ac_get_arg(&ctx->ac, ctx->rel_auto_id); LLVMValueRef vertex_dw_stride = get_tcs_in_vertex_dw_stride(ctx); @@ -3498,7 +2757,7 @@ static void si_llvm_emit_es_epilogue(struct ac_shader_abi *abi, { struct si_shader_context *ctx = si_shader_context_from_abi(abi); struct si_shader *es = ctx->shader; - struct tgsi_shader_info *info = &es->selector->info; + struct si_shader_info *info = &es->selector->info; LLVMValueRef lds_base = NULL; unsigned chan; int i; @@ -3582,25 +2841,19 @@ static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi, LLVMValueRef *addrs) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info UNUSED *info = &ctx->shader->selector->info; + struct si_shader_info UNUSED *info = &ctx->shader->selector->info; assert(info->num_outputs <= max_outputs); emit_gs_epilogue(ctx); } -static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - emit_gs_epilogue(ctx); -} - static void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi, unsigned max_outputs, LLVMValueRef *addrs) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info *info = &ctx->shader->selector->info; + struct si_shader_info *info = &ctx->shader->selector->info; struct si_shader_output_values *outputs = NULL; int i,j; @@ -3649,7 +2902,7 @@ static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi *abi, LLVMValueRef *addrs) { struct si_shader_context *ctx = si_shader_context_from_abi(abi); - struct tgsi_shader_info *info = &ctx->shader->selector->info; + struct si_shader_info *info = &ctx->shader->selector->info; LLVMValueRef pos[4] = {}; assert(info->num_outputs <= max_outputs); @@ -3671,24 +2924,15 @@ static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi *abi, ctx->return_value = ret; } -static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context *bld_base) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - ctx->abi.emit_outputs(&ctx->abi, RADEON_LLVM_MAX_OUTPUTS, - &ctx->outputs[0][0]); -} - struct si_ps_exports { unsigned num; struct ac_export_args args[10]; }; -static void si_export_mrt_z(struct lp_build_tgsi_context *bld_base, +static void si_export_mrt_z(struct si_shader_context *ctx, LLVMValueRef depth, LLVMValueRef stencil, LLVMValueRef samplemask, struct si_ps_exports *exp) { - struct si_shader_context *ctx = si_shader_context(bld_base); struct ac_export_args args; ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args); @@ -3696,12 +2940,11 @@ static void si_export_mrt_z(struct lp_build_tgsi_context *bld_base, memcpy(&exp->args[exp->num++], &args, sizeof(args)); } -static void si_export_mrt_color(struct lp_build_tgsi_context *bld_base, +static void si_export_mrt_color(struct si_shader_context *ctx, LLVMValueRef *color, unsigned index, unsigned samplemask_param, bool is_last, struct si_ps_exports *exp) { - struct si_shader_context *ctx = si_shader_context(bld_base); int i; /* Clamp color */ @@ -3716,11 +2959,11 @@ static void si_export_mrt_color(struct lp_build_tgsi_context *bld_base, /* Alpha test */ if (index == 0 && ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS) - si_alpha_test(bld_base, color[3]); + si_alpha_test(ctx, color[3]); /* Line & polygon smoothing */ if (ctx->shader->key.part.ps.epilog.poly_line_smoothing) - color[3] = si_scale_alpha_by_sample_mask(bld_base, color[3], + color[3] = si_scale_alpha_by_sample_mask(ctx, color[3], samplemask_param); /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */ @@ -3788,7 +3031,7 @@ static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi, { struct si_shader_context *ctx = si_shader_context_from_abi(abi); struct si_shader *shader = ctx->shader; - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; LLVMBuilderRef builder = ctx->ac.builder; unsigned i, j, first_vgpr, vgpr; @@ -3815,11 +3058,11 @@ static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi, break; case TGSI_SEMANTIC_POSITION: depth = LLVMBuildLoad(builder, - addrs[4 * i + 2], ""); + addrs[4 * i + 0], ""); break; case TGSI_SEMANTIC_STENCIL: stencil = LLVMBuildLoad(builder, - addrs[4 * i + 1], ""); + addrs[4 * i + 0], ""); break; case TGSI_SEMANTIC_SAMPLEMASK: samplemask = LLVMBuildLoad(builder, @@ -3867,345 +3110,6 @@ static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi, ctx->return_value = ret; } -static void membar_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0); - unsigned flags = LLVMConstIntGetZExtValue(src0); - unsigned wait_flags = 0; - - if (flags & TGSI_MEMBAR_THREAD_GROUP) - wait_flags |= AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE; - - if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER | - TGSI_MEMBAR_SHADER_BUFFER | - TGSI_MEMBAR_SHADER_IMAGE)) - wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE; - - if (flags & TGSI_MEMBAR_SHARED) - wait_flags |= AC_WAIT_LGKM; - - ac_build_waitcnt(&ctx->ac, wait_flags); -} - -static void clock_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMValueRef tmp = ac_build_shader_clock(&ctx->ac); - - emit_data->output[0] = - LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->i32_0, ""); - emit_data->output[1] = - LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->i32_1, ""); -} - -static void si_llvm_emit_ddxy( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - unsigned opcode = emit_data->info->opcode; - LLVMValueRef val; - int idx; - unsigned mask; - - if (opcode == TGSI_OPCODE_DDX_FINE) - mask = AC_TID_MASK_LEFT; - else if (opcode == TGSI_OPCODE_DDY_FINE) - mask = AC_TID_MASK_TOP; - else - mask = AC_TID_MASK_TOP_LEFT; - - /* for DDX we want to next X pixel, DDY next Y pixel. */ - idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2; - - val = ac_to_integer(&ctx->ac, emit_data->args[0]); - val = ac_build_ddxy(&ctx->ac, mask, idx, val); - emit_data->output[emit_data->chan] = val; -} - -static void build_interp_intrinsic(const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - struct si_shader *shader = ctx->shader; - const struct tgsi_shader_info *info = &shader->selector->info; - LLVMValueRef interp_param; - const struct tgsi_full_instruction *inst = emit_data->inst; - const struct tgsi_full_src_register *input = &inst->Src[0]; - int input_base, input_array_size; - int chan; - int i; - LLVMValueRef prim_mask = ac_get_arg(&ctx->ac, ctx->args.prim_mask); - LLVMValueRef array_idx, offset_x = NULL, offset_y = NULL; - int interp_param_idx; - unsigned interp; - unsigned location; - - if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) { - /* offset is in second src, first two channels */ - offset_x = lp_build_emit_fetch(bld_base, emit_data->inst, 1, - TGSI_CHAN_X); - offset_y = lp_build_emit_fetch(bld_base, emit_data->inst, 1, - TGSI_CHAN_Y); - } else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) { - LLVMValueRef sample_position; - LLVMValueRef sample_id; - LLVMValueRef halfval = LLVMConstReal(ctx->f32, 0.5f); - - /* fetch sample ID, then fetch its sample position, - * and place into first two channels. - */ - sample_id = lp_build_emit_fetch(bld_base, - emit_data->inst, 1, TGSI_CHAN_X); - sample_id = ac_to_integer(&ctx->ac, sample_id); - - /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading - * Language 4.50 spec says about interpolateAtSample: - * - * "Returns the value of the input interpolant variable at - * the location of sample number sample. If multisample - * buffers are not available, the input variable will be - * evaluated at the center of the pixel. If sample sample - * does not exist, the position used to interpolate the - * input variable is undefined." - * - * This means that sample_id values outside of the valid are - * in fact valid input, and the usual mechanism for loading the - * sample position doesn't work. - */ - if (ctx->shader->key.mono.u.ps.interpolate_at_sample_force_center) { - LLVMValueRef center[4] = { - LLVMConstReal(ctx->f32, 0.5), - LLVMConstReal(ctx->f32, 0.5), - ctx->ac.f32_0, - ctx->ac.f32_0, - }; - - sample_position = ac_build_gather_values(&ctx->ac, center, 4); - } else { - sample_position = load_sample_position(&ctx->abi, sample_id); - } - - offset_x = LLVMBuildExtractElement(ctx->ac.builder, sample_position, - ctx->i32_0, ""); - - offset_x = LLVMBuildFSub(ctx->ac.builder, offset_x, halfval, ""); - offset_y = LLVMBuildExtractElement(ctx->ac.builder, sample_position, - ctx->i32_1, ""); - offset_y = LLVMBuildFSub(ctx->ac.builder, offset_y, halfval, ""); - } - - assert(input->Register.File == TGSI_FILE_INPUT); - - if (input->Register.Indirect) { - unsigned array_id = input->Indirect.ArrayID; - - if (array_id) { - input_base = info->input_array_first[array_id]; - input_array_size = info->input_array_last[array_id] - input_base + 1; - } else { - input_base = inst->Src[0].Register.Index; - input_array_size = info->num_inputs - input_base; - } - - array_idx = si_get_indirect_index(ctx, &input->Indirect, - 1, input->Register.Index - input_base); - } else { - input_base = inst->Src[0].Register.Index; - input_array_size = 1; - array_idx = ctx->i32_0; - } - - interp = shader->selector->info.input_interpolate[input_base]; - - if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET || - inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) - location = TGSI_INTERPOLATE_LOC_CENTER; - else - location = TGSI_INTERPOLATE_LOC_CENTROID; - - interp_param_idx = lookup_interp_param_index(interp, location); - if (interp_param_idx == -1) - return; - else if (interp_param_idx) - interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx); - else - interp_param = NULL; - - if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET || - inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) { - LLVMValueRef ij_out[2]; - LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param); - - /* - * take the I then J parameters, and the DDX/Y for it, and - * calculate the IJ inputs for the interpolator. - * temp1 = ddx * offset/sample.x + I; - * interp_param.I = ddy * offset/sample.y + temp1; - * temp1 = ddx * offset/sample.x + J; - * interp_param.J = ddy * offset/sample.y + temp1; - */ - for (i = 0; i < 2; i++) { - LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, 0); - LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, 0); - LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder, - ddxy_out, ix_ll, ""); - LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder, - ddxy_out, iy_ll, ""); - LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder, - interp_param, ix_ll, ""); - LLVMValueRef temp; - - interp_el = ac_to_float(&ctx->ac, interp_el); - - temp = ac_build_fmad(&ctx->ac, ddx_el, offset_x, interp_el); - ij_out[i] = ac_build_fmad(&ctx->ac, ddy_el, offset_y, temp); - } - interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2); - } - - if (interp_param) - interp_param = ac_to_float(&ctx->ac, interp_param); - - for (chan = 0; chan < 4; chan++) { - LLVMValueRef gather = LLVMGetUndef(LLVMVectorType(ctx->f32, input_array_size)); - unsigned schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan); - - for (unsigned idx = 0; idx < input_array_size; ++idx) { - LLVMValueRef v, i = NULL, j = NULL; - - if (interp_param) { - i = LLVMBuildExtractElement( - ctx->ac.builder, interp_param, ctx->i32_0, ""); - j = LLVMBuildExtractElement( - ctx->ac.builder, interp_param, ctx->i32_1, ""); - } - v = si_build_fs_interp(ctx, input_base + idx, schan, - prim_mask, i, j); - - gather = LLVMBuildInsertElement(ctx->ac.builder, - gather, v, LLVMConstInt(ctx->i32, idx, false), ""); - } - - emit_data->output[chan] = LLVMBuildExtractElement( - ctx->ac.builder, gather, array_idx, ""); - } -} - -static void vote_all_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, emit_data->args[0]); - emit_data->output[emit_data->chan] = - LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, ""); -} - -static void vote_any_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, emit_data->args[0]); - emit_data->output[emit_data->chan] = - LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, ""); -} - -static void vote_eq_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - LLVMValueRef tmp = ac_build_vote_eq(&ctx->ac, emit_data->args[0]); - emit_data->output[emit_data->chan] = - LLVMBuildSExt(ctx->ac.builder, tmp, ctx->i32, ""); -} - -static void ballot_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - LLVMBuilderRef builder = ctx->ac.builder; - LLVMValueRef tmp; - - tmp = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X); - tmp = ac_build_ballot(&ctx->ac, tmp); - - emit_data->output[0] = LLVMBuildTrunc(builder, tmp, ctx->i32, ""); - - if (ctx->ac.wave_size == 32) { - emit_data->output[1] = ctx->i32_0; - } else { - tmp = LLVMBuildLShr(builder, tmp, LLVMConstInt(ctx->i64, 32, 0), ""); - emit_data->output[1] = LLVMBuildTrunc(builder, tmp, ctx->i32, ""); - } -} - -static void read_lane_emit( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_READ_INVOC) { - emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst, - 0, emit_data->src_chan); - - /* Always read the source invocation (= lane) from the X channel. */ - emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst, - 1, TGSI_CHAN_X); - emit_data->arg_count = 2; - } - - /* We currently have no other way to prevent LLVM from lifting the icmp - * calls to a dominating basic block. - */ - ac_build_optimization_barrier(&ctx->ac, &emit_data->args[0]); - - for (unsigned i = 0; i < emit_data->arg_count; ++i) - emit_data->args[i] = ac_to_integer(&ctx->ac, emit_data->args[i]); - - emit_data->output[emit_data->chan] = - ac_build_intrinsic(&ctx->ac, action->intr_name, - ctx->i32, emit_data->args, emit_data->arg_count, - AC_FUNC_ATTR_READNONE | - AC_FUNC_ATTR_CONVERGENT); -} - -static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - struct tgsi_src_register src0 = emit_data->inst->Src[0].Register; - LLVMValueRef imm; - unsigned stream; - - assert(src0.File == TGSI_FILE_IMMEDIATE); - - imm = ctx->imms[src0.Index * TGSI_NUM_CHANNELS + src0.SwizzleX]; - stream = LLVMConstIntGetZExtValue(imm) & 0x3; - return stream; -} - /* Emit one vertex from the geometry shader */ static void si_llvm_emit_vertex(struct ac_shader_abi *abi, unsigned stream, @@ -4218,7 +3122,7 @@ static void si_llvm_emit_vertex(struct ac_shader_abi *abi, return; } - struct tgsi_shader_info *info = &ctx->shader->selector->info; + struct si_shader_info *info = &ctx->shader->selector->info; struct si_shader *shader = ctx->shader; LLVMValueRef soffset = ac_get_arg(&ctx->ac, ctx->gs2vs_offset); LLVMValueRef gs_next_vertex; @@ -4290,18 +3194,6 @@ static void si_llvm_emit_vertex(struct ac_shader_abi *abi, ac_build_endif(&ctx->ac, 6505); } -/* Emit one vertex from the geometry shader */ -static void si_tgsi_emit_vertex( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - unsigned stream = si_llvm_get_stream(bld_base, emit_data); - - si_llvm_emit_vertex(&ctx->abi, stream, ctx->outputs[0]); -} - /* Cut one primitive from the geometry shader */ static void si_llvm_emit_primitive(struct ac_shader_abi *abi, unsigned stream) @@ -4318,23 +3210,8 @@ static void si_llvm_emit_primitive(struct ac_shader_abi *abi, si_get_gs_wave_id(ctx)); } -/* Cut one primitive from the geometry shader */ -static void si_tgsi_emit_primitive( - const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) -{ - struct si_shader_context *ctx = si_shader_context(bld_base); - - si_llvm_emit_primitive(&ctx->abi, si_llvm_get_stream(bld_base, emit_data)); -} - -static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action, - struct lp_build_tgsi_context *bld_base, - struct lp_build_emit_data *emit_data) +static void si_llvm_emit_barrier(struct si_shader_context *ctx) { - struct si_shader_context *ctx = si_shader_context(bld_base); - /* GFX6 only (thanks to a hw bug workaround): * The real barrier instruction isn’t needed, because an entire patch * always fits into a single wave. @@ -4348,27 +3225,6 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action, ac_build_s_barrier(&ctx->ac); } -void si_create_function(struct si_shader_context *ctx, - const char *name, - LLVMTypeRef *returns, unsigned num_returns, - unsigned max_workgroup_size) -{ - si_llvm_create_func(ctx, name, returns, num_returns); - ctx->return_value = LLVMGetUndef(ctx->return_type); - - if (ctx->screen->info.address32_hi) { - ac_llvm_add_target_dep_function_attr(ctx->main_fn, - "amdgpu-32bit-address-high-bits", - ctx->screen->info.address32_hi); - } - - LLVMAddTargetDependentFunctionAttr(ctx->main_fn, - "no-signed-zeros-fp-math", - "true"); - - ac_llvm_set_workgroup_size(ctx->main_fn, max_workgroup_size); -} - static void declare_streamout_params(struct si_shader_context *ctx, struct pipe_stream_output_info *so) { @@ -4481,6 +3337,28 @@ static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx) } } +static void declare_vb_descriptor_input_sgprs(struct si_shader_context *ctx) +{ + ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->vertex_buffers); + + unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs; + if (num_vbos_in_user_sgprs) { + unsigned user_sgprs = ctx->args.num_sgprs_used; + + if (is_merged_shader(ctx)) + user_sgprs -= 8; + assert(user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST); + + /* Declare unused SGPRs to align VB descriptors to 4 SGPRs (hw requirement). */ + for (unsigned i = user_sgprs; i < SI_SGPR_VS_VB_DESCRIPTOR_FIRST; i++) + ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */ + + assert(num_vbos_in_user_sgprs <= ARRAY_SIZE(ctx->vb_descriptors)); + for (unsigned i = 0; i < num_vbos_in_user_sgprs; i++) + ac_add_arg(&ctx->args, AC_ARG_SGPR, 4, AC_ARG_INT, &ctx->vb_descriptors[i]); + } +} + static void declare_vs_input_vgprs(struct si_shader_context *ctx, unsigned *num_prolog_vgprs) { @@ -4509,9 +3387,12 @@ static void declare_vs_input_vgprs(struct si_shader_context *ctx, if (!shader->is_gs_copy_shader) { /* Vertex load indices. */ - ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vertex_index0); - for (unsigned i = 1; i < shader->selector->info.num_inputs; i++) - ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); + if (shader->selector->info.num_inputs) { + ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, + &ctx->vertex_index0); + for (unsigned i = 1; i < shader->selector->info.num_inputs; i++) + ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); + } *num_prolog_vgprs += shader->selector->info.num_inputs; } } @@ -4566,7 +3447,7 @@ static void add_arg_checked(struct ac_shader_args *args, static void create_function(struct si_shader_context *ctx) { struct si_shader *shader = ctx->shader; - LLVMTypeRef returns[16+32*4]; + LLVMTypeRef returns[AC_MAX_ARGS]; unsigned i, num_return_sgprs; unsigned num_returns = 0; unsigned num_prolog_vgprs = 0; @@ -4598,10 +3479,8 @@ static void create_function(struct si_shader_context *ctx) declare_per_stage_desc_pointers(ctx, true); declare_vs_specific_input_sgprs(ctx); - if (!shader->is_gs_copy_shader) { - ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, - &ctx->vertex_buffers); - } + if (!shader->is_gs_copy_shader) + declare_vb_descriptor_input_sgprs(ctx); if (shader->key.as_es) { ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, @@ -4666,7 +3545,7 @@ static void create_function(struct si_shader_context *ctx) ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout); ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets); ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout); - ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->vertex_buffers); + declare_vb_descriptor_input_sgprs(ctx); /* VGPRs (first TCS, then VS) */ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id); @@ -4730,10 +3609,8 @@ static void create_function(struct si_shader_context *ctx) /* Declare as many input SGPRs as the VS has. */ } - if (ctx->type == PIPE_SHADER_VERTEX) { - ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, - &ctx->vertex_buffers); - } + if (ctx->type == PIPE_SHADER_VERTEX) + declare_vb_descriptor_input_sgprs(ctx); /* VGPRs (first GS, then VS/TES) */ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx01_offset); @@ -4835,12 +3712,12 @@ static void create_function(struct si_shader_context *ctx) &ctx->args.frag_pos[2], SI_PARAM_POS_Z_FLOAT); add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.frag_pos[3], SI_PARAM_POS_W_FLOAT); + shader->info.face_vgpr_index = ctx->args.num_vgprs_used; add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.front_face, SI_PARAM_FRONT_FACE); - shader->info.face_vgpr_index = 20; + shader->info.ancillary_vgpr_index = ctx->args.num_vgprs_used; add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.ancillary, SI_PARAM_ANCILLARY); - shader->info.ancillary_vgpr_index = 21; add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.sample_coverage, SI_PARAM_SAMPLE_COVERAGE); add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, @@ -4913,8 +3790,8 @@ static void create_function(struct si_shader_context *ctx) return; } - si_create_function(ctx, "main", returns, num_returns, - si_get_max_workgroup_size(shader)); + si_llvm_create_func(ctx, "main", returns, num_returns, + si_get_max_workgroup_size(shader)); /* Reserve register locations for VGPR inputs the PS prolog may need. */ if (ctx->type == PIPE_SHADER_FRAGMENT && !ctx->shader->is_monolithic) { @@ -5084,7 +3961,7 @@ static void preload_ring_buffers(struct si_shader_context *ctx) if (ctx->ac.chip_class >= GFX10) { rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) | - S_008F0C_OOB_SELECT(2) | + S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1); } else { rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | @@ -5630,7 +4507,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen, struct si_shader *shader; LLVMBuilderRef builder; struct si_shader_output_values outputs[SI_MAX_VS_OUTPUTS]; - struct tgsi_shader_info *gsinfo = &gs_selector->info; + struct si_shader_info *gsinfo = &gs_selector->info; int i; @@ -5645,9 +4522,9 @@ si_generate_gs_copy_shader(struct si_screen *sscreen, shader->selector = gs_selector; shader->is_gs_copy_shader = true; - si_init_shader_ctx(&ctx, sscreen, compiler, - si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false), - false); + si_llvm_context_init(&ctx, sscreen, compiler, + si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false), + 64); ctx.shader = shader; ctx.type = PIPE_SHADER_VERTEX; @@ -5908,51 +4785,10 @@ static void si_dump_shader_key(const struct si_shader *shader, FILE *f) } } -static void si_init_shader_ctx(struct si_shader_context *ctx, - struct si_screen *sscreen, - struct ac_llvm_compiler *compiler, - unsigned wave_size, - bool nir) -{ - struct lp_build_tgsi_context *bld_base; - - si_llvm_context_init(ctx, sscreen, compiler, wave_size, - nir ? 64 : wave_size); - - bld_base = &ctx->bld_base; - bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant; - - bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID].emit = build_interp_intrinsic; - bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE].emit = build_interp_intrinsic; - bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET].emit = build_interp_intrinsic; - - bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit; - - bld_base->op_actions[TGSI_OPCODE_CLOCK].emit = clock_emit; - - bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy; - bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy; - bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy; - bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy; - - bld_base->op_actions[TGSI_OPCODE_VOTE_ALL].emit = vote_all_emit; - bld_base->op_actions[TGSI_OPCODE_VOTE_ANY].emit = vote_any_emit; - bld_base->op_actions[TGSI_OPCODE_VOTE_EQ].emit = vote_eq_emit; - bld_base->op_actions[TGSI_OPCODE_BALLOT].emit = ballot_emit; - bld_base->op_actions[TGSI_OPCODE_READ_FIRST].intr_name = "llvm.amdgcn.readfirstlane"; - bld_base->op_actions[TGSI_OPCODE_READ_FIRST].emit = read_lane_emit; - bld_base->op_actions[TGSI_OPCODE_READ_INVOC].intr_name = "llvm.amdgcn.readlane"; - bld_base->op_actions[TGSI_OPCODE_READ_INVOC].emit = read_lane_emit; - - bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_tgsi_emit_vertex; - bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_tgsi_emit_primitive; - bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier; -} - static void si_optimize_vs_outputs(struct si_shader_context *ctx) { struct si_shader *shader = ctx->shader; - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; if ((ctx->type != PIPE_SHADER_VERTEX && ctx->type != PIPE_SHADER_TESS_EVAL) || @@ -5984,20 +4820,55 @@ static bool si_vs_needs_prolog(const struct si_shader_selector *sel, { /* VGPR initialization fixup for Vega10 and Raven is always done in the * VS prolog. */ - return sel->vs_needs_prolog || key->ls_vgpr_fix; + return sel->vs_needs_prolog || + key->ls_vgpr_fix || + key->unpack_instance_id_from_vertex_id; +} + +LLVMValueRef si_is_es_thread(struct si_shader_context *ctx) +{ + /* Return true if the current thread should execute an ES thread. */ + return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, + ac_get_thread_id(&ctx->ac), + si_unpack_param(ctx, ctx->merged_wave_info, 0, 8), ""); +} + +LLVMValueRef si_is_gs_thread(struct si_shader_context *ctx) +{ + /* Return true if the current thread should execute a GS thread. */ + return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, + ac_get_thread_id(&ctx->ac), + si_unpack_param(ctx, ctx->merged_wave_info, 8, 8), ""); +} + +static void si_llvm_emit_kill(struct ac_shader_abi *abi, LLVMValueRef visible) +{ + struct si_shader_context *ctx = si_shader_context_from_abi(abi); + LLVMBuilderRef builder = ctx->ac.builder; + + if (ctx->shader->selector->force_correct_derivs_after_kill) { + /* Kill immediately while maintaining WQM. */ + ac_build_kill_if_false(&ctx->ac, + ac_build_wqm_vote(&ctx->ac, visible)); + + LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, ""); + mask = LLVMBuildAnd(builder, mask, visible, ""); + LLVMBuildStore(builder, mask, ctx->postponed_kill); + return; + } + + ac_build_kill_if_false(&ctx->ac, visible); } -static bool si_compile_tgsi_main(struct si_shader_context *ctx, - struct nir_shader *nir, bool free_nir) +static bool si_build_main_function(struct si_shader_context *ctx, + struct nir_shader *nir, bool free_nir) { struct si_shader *shader = ctx->shader; struct si_shader_selector *sel = shader->selector; - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; // TODO clean all this up! switch (ctx->type) { case PIPE_SHADER_VERTEX: - ctx->load_input = declare_input_vs; if (shader->key.as_ls) ctx->abi.emit_outputs = si_llvm_emit_ls_epilogue; else if (shader->key.as_es) @@ -6008,22 +4879,16 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue; else ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue; - bld_base->emit_epilogue = si_tgsi_emit_epilogue; ctx->abi.load_base_vertex = get_base_vertex; break; case PIPE_SHADER_TESS_CTRL: - bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs; ctx->abi.load_tess_varyings = si_nir_load_tcs_varyings; ctx->abi.load_tess_level = si_load_tess_level; - bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs; - bld_base->emit_store = store_output_tcs; ctx->abi.store_tcs_outputs = si_nir_store_output_tcs; ctx->abi.emit_outputs = si_llvm_emit_tcs_epilogue; ctx->abi.load_patch_vertices_in = si_load_patch_vertices_in; - bld_base->emit_epilogue = si_tgsi_emit_epilogue; break; case PIPE_SHADER_TESS_EVAL: - bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes; ctx->abi.load_tess_varyings = si_nir_load_input_tes; ctx->abi.load_tess_coord = si_load_tess_coord; ctx->abi.load_tess_level = si_load_tess_level; @@ -6034,20 +4899,15 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue; else ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue; - bld_base->emit_epilogue = si_tgsi_emit_epilogue; break; case PIPE_SHADER_GEOMETRY: - bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs; ctx->abi.load_inputs = si_nir_load_input_gs; ctx->abi.emit_vertex = si_llvm_emit_vertex; ctx->abi.emit_primitive = si_llvm_emit_primitive; ctx->abi.emit_outputs = si_llvm_emit_gs_epilogue; - bld_base->emit_epilogue = si_tgsi_emit_gs_epilogue; break; case PIPE_SHADER_FRAGMENT: - ctx->load_input = declare_input_fs; ctx->abi.emit_outputs = si_llvm_return_fs_outputs; - bld_base->emit_epilogue = si_tgsi_emit_epilogue; ctx->abi.load_sample_position = load_sample_position; ctx->abi.load_sample_mask_in = load_sample_mask_in; ctx->abi.emit_fbfetch = si_nir_emit_fbfetch; @@ -6068,7 +4928,7 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, preload_ring_buffers(ctx); if (ctx->type == PIPE_SHADER_TESS_CTRL && - sel->tcs_info.tessfactors_are_def_in_all_invocs) { + sel->info.tessfactors_are_def_in_all_invocs) { for (unsigned i = 0; i < 6; i++) { ctx->invoc0_tess_factors[i] = ac_build_alloca_undef(&ctx->ac, ctx->i32, ""); @@ -6114,16 +4974,19 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, * Add an extra dword per vertex to ensure an odd stride, which * avoids bank conflicts for SoA accesses. */ - declare_esgs_ring(ctx); + if (!gfx10_is_ngg_passthrough(shader)) + declare_esgs_ring(ctx); /* This is really only needed when streamout and / or vertex * compaction is enabled. */ - LLVMTypeRef asi32 = LLVMArrayType(ctx->i32, 8); - ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module, - asi32, "ngg_scratch", AC_ADDR_SPACE_LDS); - LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32)); - LLVMSetAlignment(ctx->gs_ngg_scratch, 4); + if (sel->so.num_outputs && !ctx->gs_ngg_scratch) { + LLVMTypeRef asi32 = LLVMArrayType(ctx->i32, 8); + ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module, + asi32, "ngg_scratch", AC_ADDR_SPACE_LDS); + LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32)); + LLVMSetAlignment(ctx->gs_ngg_scratch, 4); + } } /* For GFX9 merged shaders: @@ -6144,7 +5007,6 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, */ if (ctx->screen->info.chip_class >= GFX9) { if (!shader->is_monolithic && - sel->info.num_instructions > 1 && /* not empty shader */ (shader->key.as_es || shader->key.as_ls) && (ctx->type == PIPE_SHADER_TESS_EVAL || (ctx->type == PIPE_SHADER_VERTEX && @@ -6154,7 +5016,7 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, } else if (ctx->type == PIPE_SHADER_TESS_CTRL || ctx->type == PIPE_SHADER_GEOMETRY || (shader->key.as_ngg && !shader->key.as_es)) { - LLVMValueRef num_threads; + LLVMValueRef thread_enabled; bool nested_barrier; if (!shader->is_monolithic || @@ -6171,21 +5033,15 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, nested_barrier = true; } - /* Number of patches / primitives */ - num_threads = si_unpack_param(ctx, ctx->merged_wave_info, 8, 8); + thread_enabled = si_is_gs_thread(ctx); } else { - /* Number of vertices */ - num_threads = si_unpack_param(ctx, ctx->merged_wave_info, 0, 8); + thread_enabled = si_is_es_thread(ctx); nested_barrier = false; } - LLVMValueRef ena = - LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, - ac_get_thread_id(&ctx->ac), num_threads, ""); - ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder); ctx->merged_wrap_if_label = 11500; - ac_build_ifcc(&ctx->ac, ena, ctx->merged_wrap_if_label); + ac_build_ifcc(&ctx->ac, thread_enabled, ctx->merged_wrap_if_label); if (nested_barrier) { /* Execute a barrier before the second shader in @@ -6205,7 +5061,7 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, * and contains a barrier, it will wait there and then * reach s_endpgm. */ - si_llvm_emit_barrier(NULL, bld_base, NULL); + si_llvm_emit_barrier(ctx); } } } @@ -6217,19 +5073,12 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, ctx->postponed_kill); } - if (sel->tokens) { - if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) { - fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n"); - return false; - } - } else { - bool success = si_nir_build_llvm(ctx, nir); - if (free_nir) - ralloc_free(nir); - if (!success) { - fprintf(stderr, "Failed to translate shader from NIR to LLVM\n"); - return false; - } + bool success = si_nir_build_llvm(ctx, nir); + if (free_nir) + ralloc_free(nir); + if (!success) { + fprintf(stderr, "Failed to translate shader from NIR to LLVM\n"); + return false; } si_llvm_build_ret(ctx, ctx->return_value); @@ -6246,7 +5095,7 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS. * \param key Output shader part key. */ -static void si_get_vs_prolog_key(const struct tgsi_shader_info *info, +static void si_get_vs_prolog_key(const struct si_shader_info *info, unsigned num_input_sgprs, const struct si_vs_prolog_bits *prolog_key, struct si_shader *shader_out, @@ -6255,7 +5104,7 @@ static void si_get_vs_prolog_key(const struct tgsi_shader_info *info, memset(key, 0, sizeof(*key)); key->vs_prolog.states = *prolog_key; key->vs_prolog.num_input_sgprs = num_input_sgprs; - key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1; + key->vs_prolog.num_inputs = info->num_inputs; key->vs_prolog.as_ls = shader_out->key.as_ls; key->vs_prolog.as_es = shader_out->key.as_es; key->vs_prolog.as_ngg = shader_out->key.as_ngg; @@ -6286,7 +5135,7 @@ static void si_get_ps_prolog_key(struct si_shader *shader, union si_shader_part_key *key, bool separate_prolog) { - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; memset(key, 0, sizeof(*key)); key->ps_prolog.states = shader->key.part.ps.prolog; @@ -6435,7 +5284,7 @@ static bool si_need_ps_prolog(const union si_shader_part_key *key) static void si_get_ps_epilog_key(struct si_shader *shader, union si_shader_part_key *key) { - struct tgsi_shader_info *info = &shader->selector->info; + struct si_shader_info *info = &shader->selector->info; memset(key, 0, sizeof(*key)); key->ps_epilog.colors_written = info->colors_written; key->ps_epilog.writes_z = info->writes_z; @@ -6453,7 +5302,7 @@ static void si_build_gs_prolog_function(struct si_shader_context *ctx, { unsigned num_sgprs, num_vgprs; LLVMBuilderRef builder = ctx->ac.builder; - LLVMTypeRef returns[48]; + LLVMTypeRef returns[AC_MAX_ARGS]; LLVMValueRef func, ret; memset(&ctx->args, 0, sizeof(ctx->args)); @@ -6480,8 +5329,7 @@ static void si_build_gs_prolog_function(struct si_shader_context *ctx, } /* Create the function. */ - si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs, - 0); + si_llvm_create_func(ctx, "gs_prolog", returns, num_sgprs + num_vgprs, 0); func = ctx->main_fn; /* Set the full EXEC mask for the prolog, because we are only fiddling @@ -6580,9 +5428,9 @@ static void si_build_wrapper_function(struct si_shader_context *ctx, { LLVMBuilderRef builder = ctx->ac.builder; /* PS epilog has one arg per color component; gfx9 merged shader - * prologs need to forward 32 user SGPRs. + * prologs need to forward 40 SGPRs. */ - LLVMValueRef initial[64], out[64]; + LLVMValueRef initial[AC_MAX_ARGS], out[AC_MAX_ARGS]; LLVMTypeRef function_type; unsigned num_first_params; unsigned num_out, initial_num_out; @@ -6648,7 +5496,7 @@ static void si_build_wrapper_function(struct si_shader_context *ctx, /* Prepare the return type. */ unsigned num_returns = 0; - LLVMTypeRef returns[32], last_func_type, return_type; + LLVMTypeRef returns[AC_MAX_ARGS], last_func_type, return_type; last_func_type = LLVMGetElementType(LLVMTypeOf(parts[num_parts - 1])); return_type = LLVMGetReturnType(last_func_type); @@ -6665,8 +5513,8 @@ static void si_build_wrapper_function(struct si_shader_context *ctx, unreachable("unexpected type"); } - si_create_function(ctx, "wrapper", returns, num_returns, - si_get_max_workgroup_size(ctx->shader)); + si_llvm_create_func(ctx, "wrapper", returns, num_returns, + si_get_max_workgroup_size(ctx->shader)); if (is_merged_shader(ctx)) ac_init_exec_full_mask(&ctx->ac); @@ -6719,7 +5567,7 @@ static void si_build_wrapper_function(struct si_shader_context *ctx, /* Now chain the parts. */ LLVMValueRef ret = NULL; for (unsigned part = 0; part < num_parts; ++part) { - LLVMValueRef in[48]; + LLVMValueRef in[AC_MAX_ARGS]; LLVMTypeRef ret_type; unsigned out_idx = 0; unsigned num_params = LLVMCountParams(parts[part]); @@ -6875,10 +5723,10 @@ static struct nir_shader *get_nir_shader(struct si_shader_selector *sel, return NULL; } -int si_compile_tgsi_shader(struct si_screen *sscreen, - struct ac_llvm_compiler *compiler, - struct si_shader *shader, - struct pipe_debug_callback *debug) +int si_compile_shader(struct si_screen *sscreen, + struct ac_llvm_compiler *compiler, + struct si_shader *shader, + struct pipe_debug_callback *debug) { struct si_shader_selector *sel = shader->selector; struct si_shader_context ctx; @@ -6886,34 +5734,30 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, struct nir_shader *nir = get_nir_shader(sel, &free_nir); int r = -1; - /* Dump TGSI code before doing TGSI->LLVM conversion in case the + /* Dump NIR before doing NIR->LLVM conversion in case the * conversion fails. */ if (si_can_dump_shader(sscreen, sel->type) && - !(sscreen->debug_flags & DBG(NO_TGSI))) { - if (sel->tokens) - tgsi_dump(sel->tokens, 0); - else - nir_print_shader(nir, stderr); + !(sscreen->debug_flags & DBG(NO_NIR))) { + nir_print_shader(nir, stderr); si_dump_streamout(&sel->so); } - si_init_shader_ctx(&ctx, sscreen, compiler, si_get_shader_wave_size(shader), - nir != NULL); - si_llvm_context_set_ir(&ctx, shader, nir); + si_llvm_context_init(&ctx, sscreen, compiler, si_get_shader_wave_size(shader), 64); + si_llvm_context_set_ir(&ctx, shader); memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED, sizeof(shader->info.vs_output_param_offset)); shader->info.uses_instanceid = sel->info.uses_instanceid; - if (!si_compile_tgsi_main(&ctx, nir, free_nir)) { + if (!si_build_main_function(&ctx, nir, free_nir)) { si_llvm_dispose(&ctx); return -1; } if (shader->is_monolithic && ctx.type == PIPE_SHADER_VERTEX) { LLVMValueRef parts[2]; - bool need_prolog = sel->vs_needs_prolog; + bool need_prolog = si_vs_needs_prolog(sel, &shader->key.part.vs.prolog); parts[1] = ctx.main_fn; @@ -6923,6 +5767,7 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, shader->info.num_input_sgprs, &shader->key.part.vs.prolog, shader, &prolog_key); + prolog_key.vs_prolog.is_monolithic = true; si_build_vs_prolog_function(&ctx, &prolog_key); parts[0] = ctx.main_fn; } @@ -6957,9 +5802,9 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, shader_ls.key.mono = shader->key.mono; shader_ls.key.opt = shader->key.opt; shader_ls.is_monolithic = true; - si_llvm_context_set_ir(&ctx, &shader_ls, nir); + si_llvm_context_set_ir(&ctx, &shader_ls); - if (!si_compile_tgsi_main(&ctx, nir, free_nir)) { + if (!si_build_main_function(&ctx, nir, free_nir)) { si_llvm_dispose(&ctx); return -1; } @@ -7025,9 +5870,9 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, shader_es.key.mono = shader->key.mono; shader_es.key.opt = shader->key.opt; shader_es.is_monolithic = true; - si_llvm_context_set_ir(&ctx, &shader_es, nir); + si_llvm_context_set_ir(&ctx, &shader_es); - if (!si_compile_tgsi_main(&ctx, nir, free_nir)) { + if (!si_build_main_function(&ctx, nir, free_nir)) { si_llvm_dispose(&ctx); return -1; } @@ -7035,7 +5880,8 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, es_main = ctx.main_fn; /* ES prolog */ - if (es->vs_needs_prolog) { + if (es->type == PIPE_SHADER_VERTEX && + si_vs_needs_prolog(es, &shader->key.part.gs.vs_prolog)) { union si_shader_part_key vs_prolog_key; si_get_vs_prolog_key(&es->info, shader_es.info.num_input_sgprs, @@ -7243,10 +6089,10 @@ si_get_shader_part(struct si_screen *sscreen, } struct si_shader_context ctx; - si_init_shader_ctx(&ctx, sscreen, compiler, - si_get_wave_size(sscreen, type, shader.key.as_ngg, - shader.key.as_es), - false); + si_llvm_context_init(&ctx, sscreen, compiler, + si_get_wave_size(sscreen, type, shader.key.as_ngg, + shader.key.as_es), + 64); ctx.shader = &shader; ctx.type = type; @@ -7317,7 +6163,7 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, memset(&ctx->args, 0, sizeof(ctx->args)); /* 4 preloaded VGPRs + vertex load indices as prolog outputs */ - returns = alloca((num_all_input_regs + key->vs_prolog.last_input + 1) * + returns = alloca((num_all_input_regs + key->vs_prolog.num_inputs) * sizeof(LLVMTypeRef)); num_returns = 0; @@ -7337,11 +6183,11 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, } /* Vertex load indices. */ - for (i = 0; i <= key->vs_prolog.last_input; i++) + for (i = 0; i < key->vs_prolog.num_inputs; i++) returns[num_returns++] = ctx->f32; /* Create the function. */ - si_create_function(ctx, "vs_prolog", returns, num_returns, 0); + si_llvm_create_func(ctx, "vs_prolog", returns, num_returns, 0); func = ctx->main_fn; for (i = 0; i < num_input_vgprs; i++) { @@ -7412,22 +6258,6 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, key->vs_prolog.num_input_sgprs + i, ""); } - LLVMValueRef original_ret = ret; - bool wrapped = false; - LLVMBasicBlockRef if_entry_block = NULL; - - if (key->vs_prolog.is_monolithic && key->vs_prolog.as_ngg) { - LLVMValueRef num_threads; - LLVMValueRef ena; - - num_threads = si_unpack_param(ctx, merged_wave_info, 0, 8); - ena = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, - ac_get_thread_id(&ctx->ac), num_threads, ""); - if_entry_block = LLVMGetInsertBlock(ctx->ac.builder); - ac_build_ifcc(&ctx->ac, ena, 11501); - wrapped = true; - } - /* Compute vertex load indices from instance divisors. */ LLVMValueRef instance_divisor_constbuf = NULL; @@ -7439,7 +6269,7 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, ac_build_load_to_sgpr(&ctx->ac, list, buf_index); } - for (i = 0; i <= key->vs_prolog.last_input; i++) { + for (i = 0; i < key->vs_prolog.num_inputs; i++) { bool divisor_is_one = key->vs_prolog.states.instance_divisor_is_one & (1u << i); bool divisor_is_fetched = @@ -7483,20 +6313,6 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, ctx->args.arg_count + i, ""); } - if (wrapped) { - LLVMBasicBlockRef bbs[2] = { - LLVMGetInsertBlock(ctx->ac.builder), - if_entry_block, - }; - ac_build_endif(&ctx->ac, 11501); - - LLVMValueRef values[2] = { - ret, - original_ret - }; - ret = ac_build_phi(&ctx->ac, LLVMTypeOf(ret), 2, values, bbs); - } - si_llvm_build_ret(ctx, ret); } @@ -7544,8 +6360,6 @@ static bool si_shader_select_vs_parts(struct si_screen *sscreen, static void si_build_tcs_epilog_function(struct si_shader_context *ctx, union si_shader_part_key *key) { - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; - memset(&ctx->args, 0, sizeof(ctx->args)); if (ctx->screen->info.chip_class >= GFX9) { @@ -7604,15 +6418,15 @@ static void si_build_tcs_epilog_function(struct si_shader_context *ctx, ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &tess_factors[i]); /* Create the function. */ - si_create_function(ctx, "tcs_epilog", NULL, 0, - ctx->screen->info.chip_class >= GFX7 ? 128 : 0); + si_llvm_create_func(ctx, "tcs_epilog", NULL, 0, + ctx->screen->info.chip_class >= GFX7 ? 128 : 0); ac_declare_lds_as_pointer(&ctx->ac); LLVMValueRef invoc0_tess_factors[6]; for (unsigned i = 0; i < 6; i++) invoc0_tess_factors[i] = ac_get_arg(&ctx->ac, tess_factors[i]); - si_write_tess_factors(bld_base, + si_write_tess_factors(ctx, ac_get_arg(&ctx->ac, rel_patch_id), ac_get_arg(&ctx->ac, invocation_id), ac_get_arg(&ctx->ac, tcs_out_current_patch_data_offset), @@ -7749,7 +6563,7 @@ static void si_build_ps_prolog_function(struct si_shader_context *ctx, return_types[num_returns++] = ctx->f32; /* Create the function. */ - si_create_function(ctx, "ps_prolog", return_types, num_returns, 0); + si_llvm_create_func(ctx, "ps_prolog", return_types, num_returns, 0); func = ctx->main_fn; /* Copy inputs to outputs. This should be no-op, as the registers match, @@ -7918,9 +6732,8 @@ static void si_build_ps_prolog_function(struct si_shader_context *ctx, face = ac_to_integer(&ctx->ac, face); } - interp_fs_input(ctx, - key->ps_prolog.color_attr_index[i], - TGSI_SEMANTIC_COLOR, i, + interp_fs_color(ctx, + key->ps_prolog.color_attr_index[i], i, key->ps_prolog.num_interp_inputs, key->ps_prolog.colors_read, interp_ij, prim_mask, face, color); @@ -7994,7 +6807,6 @@ static void si_build_ps_prolog_function(struct si_shader_context *ctx, static void si_build_ps_epilog_function(struct si_shader_context *ctx, union si_shader_part_key *key) { - struct lp_build_tgsi_context *bld_base = &ctx->bld_base; LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL; int i; struct si_ps_exports exp = {}; @@ -8027,7 +6839,7 @@ static void si_build_ps_epilog_function(struct si_shader_context *ctx, ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL); /* Create the function. */ - si_create_function(ctx, "ps_epilog", NULL, 0, 0); + si_llvm_create_func(ctx, "ps_epilog", NULL, 0, 0); /* Disable elimination of unused inputs. */ ac_llvm_add_target_dep_function_attr(ctx->main_fn, "InitialPSInputAddr", 0xffffff); @@ -8064,7 +6876,7 @@ static void si_build_ps_epilog_function(struct si_shader_context *ctx, for (i = 0; i < 4; i++) color[i] = LLVMGetParam(ctx->main_fn, vgpr++); - si_export_mrt_color(bld_base, color, mrt, + si_export_mrt_color(ctx, color, mrt, ctx->args.arg_count - 1, mrt == last_color_export, &exp); } @@ -8078,7 +6890,7 @@ static void si_build_ps_epilog_function(struct si_shader_context *ctx, samplemask = LLVMGetParam(ctx->main_fn, vgpr++); if (depth || stencil || samplemask) - si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp); + si_export_mrt_z(ctx, depth, stencil, samplemask, &exp); else if (last_color_export == -1) ac_build_export_null(&ctx->ac); @@ -8223,9 +7035,10 @@ static void si_fix_resource_usage(struct si_screen *sscreen, } } -bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compiler, - struct si_shader *shader, - struct pipe_debug_callback *debug) +bool si_create_shader_variant(struct si_screen *sscreen, + struct ac_llvm_compiler *compiler, + struct si_shader *shader, + struct pipe_debug_callback *debug) { struct si_shader_selector *sel = shader->selector; struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key); @@ -8244,7 +7057,7 @@ bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compil /* Monolithic shader (compiled as a whole, has many variants, * may take a long time to compile). */ - r = si_compile_tgsi_shader(sscreen, compiler, shader, debug); + r = si_compile_shader(sscreen, compiler, shader, debug); if (r) return false; } else { @@ -8266,7 +7079,7 @@ bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compil if (!mainp) return false; - /* Copy the compiled TGSI shader data over. */ + /* Copy the compiled shader data over. */ shader->is_binary_shared = true; shader->binary = mainp->binary; shader->config = mainp->config;