* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
+#include <llvm/Config/llvm-config.h>
+
#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 "sid.h"
#include "compiler/nir/nir.h"
+#include "compiler/nir/nir_serialize.h"
static const char scratch_rsrc_dword0_symbol[] =
"SCRATCH_RSRC_DWORD0";
static const char scratch_rsrc_dword1_symbol[] =
"SCRATCH_RSRC_DWORD1";
-struct si_shader_output_values
-{
- LLVMValueRef values[4];
- unsigned semantic_name;
- unsigned semantic_index;
- ubyte vertex_stream[4];
-};
-
static void si_init_shader_ctx(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler);
+ 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_dump_shader_key(unsigned processor, const struct si_shader *shader,
- FILE *f);
+static void si_dump_shader_key(const struct si_shader *shader, FILE *f);
static void si_build_vs_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key);
return false;
}
-static bool is_merged_shader(struct si_shader_context *ctx)
+/** Whether the shader runs as a combination of multiple API shaders */
+static bool is_multi_part_shader(struct si_shader_context *ctx)
{
if (ctx->screen->info.chip_class <= GFX8)
return false;
ctx->type == PIPE_SHADER_GEOMETRY;
}
-void si_init_function_info(struct si_function_info *fninfo)
-{
- fninfo->num_params = 0;
- fninfo->num_sgpr_params = 0;
-}
-
-unsigned add_arg_assign(struct si_function_info *fninfo,
- enum si_arg_regfile regfile, LLVMTypeRef type,
- LLVMValueRef *assign)
-{
- assert(regfile != ARG_SGPR || fninfo->num_sgpr_params == fninfo->num_params);
-
- unsigned idx = fninfo->num_params++;
- assert(idx < ARRAY_SIZE(fninfo->types));
-
- if (regfile == ARG_SGPR)
- fninfo->num_sgpr_params = fninfo->num_params;
-
- fninfo->types[idx] = type;
- fninfo->assign[idx] = assign;
- return idx;
-}
-
-static unsigned add_arg(struct si_function_info *fninfo,
- enum si_arg_regfile regfile, LLVMTypeRef type)
-{
- return add_arg_assign(fninfo, regfile, type, NULL);
-}
-
-static void add_arg_assign_checked(struct si_function_info *fninfo,
- enum si_arg_regfile regfile, LLVMTypeRef type,
- LLVMValueRef *assign, unsigned idx)
-{
- MAYBE_UNUSED unsigned actual = add_arg_assign(fninfo, regfile, type, assign);
- assert(actual == idx);
-}
-
-static void add_arg_checked(struct si_function_info *fninfo,
- enum si_arg_regfile regfile, LLVMTypeRef type,
- unsigned idx)
+/** Whether the shader runs on a merged HW stage (LSHS or ESGS) */
+static bool is_merged_shader(struct si_shader_context *ctx)
{
- add_arg_assign_checked(fninfo, regfile, type, NULL, idx);
+ return ctx->shader->key.as_ngg || is_multi_part_shader(ctx);
}
/**
assert(!"invalid generic index");
return 0;
- case TGSI_SEMANTIC_PSIZE:
- return SI_MAX_IO_GENERIC + 1;
- case TGSI_SEMANTIC_CLIPDIST:
- assert(index <= 1);
- return SI_MAX_IO_GENERIC + 2 + index;
case TGSI_SEMANTIC_FOG:
- return SI_MAX_IO_GENERIC + 4;
- case TGSI_SEMANTIC_LAYER:
- return SI_MAX_IO_GENERIC + 5;
- case TGSI_SEMANTIC_VIEWPORT_INDEX:
- return SI_MAX_IO_GENERIC + 6;
- case TGSI_SEMANTIC_PRIMID:
- return SI_MAX_IO_GENERIC + 7;
+ return SI_MAX_IO_GENERIC + 1;
case TGSI_SEMANTIC_COLOR:
assert(index < 2);
- return SI_MAX_IO_GENERIC + 8 + index;
+ return SI_MAX_IO_GENERIC + 2 + index;
case TGSI_SEMANTIC_BCOLOR:
assert(index < 2);
/* If it's a varying, COLOR and BCOLOR alias. */
if (is_varying)
- return SI_MAX_IO_GENERIC + 8 + index;
+ return SI_MAX_IO_GENERIC + 2 + index;
else
- return SI_MAX_IO_GENERIC + 10 + index;
+ return SI_MAX_IO_GENERIC + 4 + index;
case TGSI_SEMANTIC_TEXCOORD:
assert(index < 8);
- STATIC_ASSERT(SI_MAX_IO_GENERIC + 12 + 8 <= 63);
- return SI_MAX_IO_GENERIC + 12 + index;
+ return SI_MAX_IO_GENERIC + 6 + index;
+
+ /* These are rarely used between LS and HS or ES and GS. */
+ case TGSI_SEMANTIC_CLIPDIST:
+ assert(index < 2);
+ return SI_MAX_IO_GENERIC + 6 + 8 + index;
case TGSI_SEMANTIC_CLIPVERTEX:
- return 63;
+ return SI_MAX_IO_GENERIC + 6 + 8 + 2;
+ case TGSI_SEMANTIC_PSIZE:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 3;
+
+ /* These can't be written by LS, HS, and ES. */
+ case TGSI_SEMANTIC_LAYER:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 4;
+ case TGSI_SEMANTIC_VIEWPORT_INDEX:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 5;
+ case TGSI_SEMANTIC_PRIMID:
+ STATIC_ASSERT(SI_MAX_IO_GENERIC + 6 + 8 + 6 <= 63);
+ return SI_MAX_IO_GENERIC + 6 + 8 + 6;
default:
fprintf(stderr, "invalid semantic name = %u\n", semantic_name);
assert(!"invalid semantic name");
}
LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
- unsigned param, unsigned rshift,
+ struct ac_arg param, unsigned rshift,
unsigned bitwidth)
{
- LLVMValueRef value = LLVMGetParam(ctx->main_fn, param);
+ LLVMValueRef value = ac_get_arg(&ctx->ac, param);
return unpack_llvm_param(ctx, value, rshift, bitwidth);
}
{
switch (ctx->type) {
case PIPE_SHADER_TESS_CTRL:
- return unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 0, 8);
+ return si_unpack_param(ctx, ctx->args.tcs_rel_ids, 0, 8);
case PIPE_SHADER_TESS_EVAL:
- return LLVMGetParam(ctx->main_fn,
- ctx->param_tes_rel_patch_id);
+ return ac_get_arg(&ctx->ac, ctx->tes_rel_patch_id);
default:
assert(0);
static LLVMValueRef
get_tcs_in_patch_stride(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, ctx->param_vs_state_bits, 8, 13);
+ return si_unpack_param(ctx, ctx->vs_state_bits, 8, 13);
}
static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context *ctx)
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->param_tcs_out_lds_layout, 0, 13);
+ return si_unpack_param(ctx, ctx->tcs_out_lds_layout, 0, 13);
const struct tgsi_shader_info *info = &ctx->shader->selector->info;
unsigned tcs_out_vertices = info->properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
get_tcs_out_patch0_offset(struct si_shader_context *ctx)
{
return LLVMBuildMul(ctx->ac.builder,
- si_unpack_param(ctx,
- ctx->param_tcs_out_lds_offsets,
- 0, 16),
+ si_unpack_param(ctx, ctx->tcs_out_lds_offsets, 0, 16),
LLVMConstInt(ctx->i32, 4, 0), "");
}
get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
{
return LLVMBuildMul(ctx->ac.builder,
- si_unpack_param(ctx,
- ctx->param_tcs_out_lds_offsets,
- 16, 16),
+ si_unpack_param(ctx, ctx->tcs_out_lds_offsets, 16, 16),
LLVMConstInt(ctx->i32, 4, 0), "");
}
if (ctx->type == PIPE_SHADER_TESS_CTRL && tcs_out_vertices)
return LLVMConstInt(ctx->i32, tcs_out_vertices, 0);
- return si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 6, 6);
+ return si_unpack_param(ctx, ctx->tcs_offchip_layout, 6, 6);
}
static LLVMValueRef get_tcs_in_vertex_dw_stride(struct si_shader_context *ctx)
stride = ctx->shader->key.part.tcs.ls->lshs_vertex_stride / 4;
return LLVMConstInt(ctx->i32, stride, 0);
}
- return si_unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
+ return si_unpack_param(ctx, ctx->vs_state_bits, 24, 8);
default:
assert(0);
LLVMValueRef out[4])
{
const struct tgsi_shader_info *info = &ctx->shader->selector->info;
- unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+ unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
if (vs_blit_property) {
LLVMValueRef vertex_id = ctx->abi.vertex_id;
LLVMIntNE, vertex_id,
ctx->i32_1, "");
+ unsigned param_vs_blit_inputs = ctx->vs_blit_inputs.arg_index;
if (input_index == 0) {
/* Position: */
LLVMValueRef x1y1 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs);
+ param_vs_blit_inputs);
LLVMValueRef x2y2 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 1);
+ param_vs_blit_inputs + 1);
LLVMValueRef x1 = unpack_sint16(ctx, x1y1, 0);
LLVMValueRef y1 = unpack_sint16(ctx, x1y1, 1);
out[0] = LLVMBuildSIToFP(ctx->ac.builder, x, ctx->f32, "");
out[1] = LLVMBuildSIToFP(ctx->ac.builder, y, ctx->f32, "");
out[2] = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 2);
+ param_vs_blit_inputs + 2);
out[3] = ctx->ac.f32_1;
return;
}
if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
for (int i = 0; i < 4; i++) {
out[i] = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 3 + i);
+ param_vs_blit_inputs + 3 + i);
}
} else {
assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD);
LLVMValueRef x1 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 3);
+ param_vs_blit_inputs + 3);
LLVMValueRef y1 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 4);
+ param_vs_blit_inputs + 4);
LLVMValueRef x2 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 5);
+ param_vs_blit_inputs + 5);
LLVMValueRef y2 = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 6);
+ param_vs_blit_inputs + 6);
out[0] = LLVMBuildSelect(ctx->ac.builder, sel_x1,
x1, x2, "");
out[1] = LLVMBuildSelect(ctx->ac.builder, sel_y1,
y1, y2, "");
out[2] = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 7);
+ param_vs_blit_inputs + 7);
out[3] = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_blit_inputs + 8);
+ param_vs_blit_inputs + 8);
}
return;
}
LLVMValueRef tmp;
/* Load the T list */
- t_list_ptr = LLVMGetParam(ctx->main_fn, ctx->param_vertex_buffers);
+ 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);
vertex_index = LLVMGetParam(ctx->main_fn,
- ctx->param_vertex_index0 +
+ ctx->vertex_index0.arg_index +
input_index);
/* Use the open-coded implementation for all loads of doubles and
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,
- false, false, true);
+ t_list, 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;
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,
- channels_per_fetch, false, true);
+ channels_per_fetch, 0, true);
}
if (num_fetches == 1 && channels_per_fetch > 1) {
si_llvm_load_input_vs(ctx, input_index, out);
}
-static LLVMValueRef get_primitive_id(struct si_shader_context *ctx,
- unsigned swizzle)
+LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx,
+ unsigned swizzle)
{
if (swizzle > 0)
return ctx->i32_0;
switch (ctx->type) {
case PIPE_SHADER_VERTEX:
- return LLVMGetParam(ctx->main_fn,
- ctx->param_vs_prim_id);
+ return ac_get_arg(&ctx->ac, ctx->vs_prim_id);
case PIPE_SHADER_TESS_CTRL:
- return ctx->abi.tcs_patch_id;
+ return ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id);
case PIPE_SHADER_TESS_EVAL:
- return ctx->abi.tes_patch_id;
+ return ac_get_arg(&ctx->ac, ctx->args.tes_patch_id);
case PIPE_SHADER_GEOMETRY:
- return ctx->abi.gs_prim_id;
+ return ac_get_arg(&ctx->ac, ctx->args.gs_prim_id);
default:
assert(0);
return ctx->i32_0;
LLVMValueRef base_addr,
LLVMValueRef vertex_index,
LLVMValueRef param_index,
- unsigned input_index,
- ubyte *name,
- ubyte *index,
- bool is_patch)
+ ubyte name, ubyte index)
{
if (vertex_dw_stride) {
base_addr = ac_build_imad(&ctx->ac, vertex_index,
LLVMConstInt(ctx->i32, 4, 0), base_addr);
}
- int param = is_patch ?
- si_shader_io_get_unique_index_patch(name[input_index],
- index[input_index]) :
- si_shader_io_get_unique_index(name[input_index],
- index[input_index], false);
+ int param = 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);
/* Add the base address of the element. */
return LLVMBuildAdd(ctx->ac.builder, base_addr,
return get_dw_address_from_generic_indices(ctx, vertex_dw_stride,
base_addr, vertex_index,
- ind_index, input_index,
- name, index,
- !reg.Register.Dimension);
+ ind_index, name[input_index],
+ index[input_index]);
}
/* The offchip buffer layout for TCS->TES is
LLVMValueRef param_stride, constant16;
vertices_per_patch = get_num_tcs_out_vertices(ctx);
- num_patches = si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 0, 6);
+ num_patches = si_unpack_param(ctx, ctx->tcs_offchip_layout, 0, 6);
total_vertices = LLVMBuildMul(ctx->ac.builder, vertices_per_patch,
num_patches, "");
if (!vertex_index) {
LLVMValueRef patch_data_offset =
- si_unpack_param(ctx, ctx->param_tcs_offchip_layout, 12, 20);
+ si_unpack_param(ctx, ctx->tcs_offchip_layout, 12, 20);
base_addr = LLVMBuildAdd(ctx->ac.builder, base_addr,
patch_data_offset, "");
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,
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);
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);
+ param_index, name[param_base],
+ index[param_base]);
}
static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
if (swizzle == ~0) {
value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
- 0, 1, 0, can_speculate, false);
+ 0, ac_glc, can_speculate, false);
return LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
}
if (!llvm_type_is_64bit(ctx, type)) {
value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
- 0, 1, 0, can_speculate, false);
+ 0, ac_glc, can_speculate, false);
value = LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
return LLVMBuildExtractElement(ctx->ac.builder, value,
}
value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
- swizzle * 4, 1, 0, can_speculate, false);
+ swizzle * 4, ac_glc, can_speculate, false);
value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
- swizzle * 4 + 4, 1, 0, can_speculate, false);
+ swizzle * 4 + 4, ac_glc, can_speculate, false);
return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
}
enum si_tess_ring ring)
{
LLVMBuilderRef builder = ctx->ac.builder;
- unsigned param = ring == TESS_OFFCHIP_RING_TES ? ctx->param_tes_offchip_addr :
- ctx->param_tcs_out_lds_layout;
- LLVMValueRef addr = LLVMGetParam(ctx->main_fn, param);
+ LLVMValueRef addr = ac_get_arg(&ctx->ac,
+ ring == TESS_OFFCHIP_RING_TES ?
+ ctx->tes_offchip_addr :
+ ctx->tcs_out_lds_layout);
/* TCS only receives high 13 bits of the address. */
if (ring == TESS_OFFCHIP_RING_TCS || ring == TCS_FACTOR_RING) {
LLVMConstInt(ctx->i32, tf_offset, 0), "");
}
+ uint32_t rsrc3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
+
+ if (ctx->screen->info.chip_class >= GFX10)
+ rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
+ S_008F0C_OOB_SELECT(3) |
+ S_008F0C_RESOURCE_LEVEL(1);
+ else
+ rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
LLVMValueRef desc[4];
desc[0] = addr;
desc[1] = LLVMConstInt(ctx->i32,
S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);
desc[2] = LLVMConstInt(ctx->i32, 0xffffffff, 0);
- desc[3] = LLVMConstInt(ctx->i32,
- S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
- S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
- S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
- S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
- S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
- S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32), 0);
+ desc[3] = LLVMConstInt(ctx->i32, rsrc3, false);
return ac_build_gather_values(&ctx->ac, desc, 4);
}
struct tgsi_shader_info *info = &ctx->shader->selector->info;
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
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);
}
}
- if (param_index) {
- /* Add the constant index to the indirect index */
- param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
- LLVMConstInt(ctx->i32, const_index, 0), "");
- } else {
+ if (!param_index) {
param_index = LLVMConstInt(ctx->i32, const_index, 0);
}
- ubyte *names;
- ubyte *indices;
- if (load_input) {
- names = info->input_semantic_name;
- indices = info->input_semantic_index;
- } else {
- names = info->output_semantic_name;
- indices = info->output_semantic_index;
- }
-
dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
vertex_index, param_index,
- driver_location,
- names, indices,
- is_patch);
+ name, index);
LLVMValueRef value[4];
for (unsigned i = 0; i < num_components; i++) {
LLVMValueRef base, addr;
unsigned swizzle = (swizzle_in & 0xffff);
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ 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,
LLVMValueRef base, addr;
driver_location = driver_location / 4;
+ ubyte name = info->input_semantic_name[driver_location];
+ ubyte index = info->input_semantic_index[driver_location];
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ assert((name == TGSI_SEMANTIC_PATCH ||
+ name == TGSI_SEMANTIC_TESSINNER ||
+ name == TGSI_SEMANTIC_TESSOUTER) == is_patch);
- if (param_index) {
- /* Add the constant index to the indirect index */
- param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
- LLVMConstInt(ctx->i32, const_index, 0), "");
- } else {
+ base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
+
+ if (!param_index) {
param_index = LLVMConstInt(ctx->i32, const_index, 0);
}
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
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;
buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ 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;
if (reg->Register.WriteMask != 0xF && !is_tess_factor) {
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
buf_addr, base,
- 4 * chan_index, 1, 0, true, false);
+ 4 * chan_index, ac_glc);
}
/* Write tess factors into VGPRs for the epilog. */
LLVMValueRef value = ac_build_gather_values(&ctx->ac,
values, 4);
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buf_addr,
- base, 0, 1, 0, true, false);
+ base, 0, ac_glc);
}
}
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
struct tgsi_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;
bool is_tess_factor = false, is_tess_inner = false;
driver_location = driver_location / 4;
+ ubyte name = info->output_semantic_name[driver_location];
+ ubyte index = info->output_semantic_index[driver_location];
- if (param_index) {
- /* Add the constant index to the indirect index */
- param_index = LLVMBuildAdd(ctx->ac.builder, param_index,
- LLVMConstInt(ctx->i32, const_index, 0), "");
- } else {
- if (const_index != 0)
- param_index = LLVMConstInt(ctx->i32, const_index, 0);
- }
+ bool is_const = !param_index;
+ 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);
dw_addr = get_dw_address_from_generic_indices(ctx, stride, dw_addr,
vertex_index, param_index,
- driver_location,
- info->output_semantic_name,
- info->output_semantic_index,
- is_patch);
+ name, index);
skip_lds_store = !info->reads_pervertex_outputs;
} else {
dw_addr = get_tcs_out_current_patch_data_offset(ctx);
dw_addr = get_dw_address_from_generic_indices(ctx, NULL, dw_addr,
vertex_index, param_index,
- driver_location,
- info->output_semantic_name,
- info->output_semantic_index,
- is_patch);
+ name, index);
skip_lds_store = !info->reads_perpatch_outputs;
- if (!param_index) {
+ if (is_const && const_index == 0) {
int name = info->output_semantic_name[driver_location];
/* Always write tess factors into LDS for the TCS epilog. */
buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ 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. */
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
addr, base,
4 * buffer_store_offset,
- 1, 0, true, false);
+ ac_glc);
}
/* Write tess factors into VGPRs for the epilog. */
LLVMValueRef value = ac_build_gather_values(&ctx->ac,
values, 4);
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, addr,
- base, 0, 1, 0, true, false);
+ base, 0, ac_glc);
}
}
switch (index / 2) {
case 0:
- vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx01_offset,
- index % 2 ? 16 : 0, 16);
+ vtx_offset = si_unpack_param(ctx, ctx->gs_vtx01_offset,
+ index % 2 ? 16 : 0, 16);
break;
case 1:
- vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx23_offset,
- index % 2 ? 16 : 0, 16);
+ vtx_offset = si_unpack_param(ctx, ctx->gs_vtx23_offset,
+ index % 2 ? 16 : 0, 16);
break;
case 2:
- vtx_offset = si_unpack_param(ctx, ctx->param_gs_vtx45_offset,
- index % 2 ? 16 : 0, 16);
+ vtx_offset = si_unpack_param(ctx, ctx->gs_vtx45_offset,
+ index % 2 ? 16 : 0, 16);
break;
default:
assert(0);
}
/* Get the vertex offset parameter on GFX6. */
- LLVMValueRef gs_vtx_offset = ctx->gs_vtx_offset[vtx_offset_param];
+ LLVMValueRef gs_vtx_offset = ac_get_arg(&ctx->ac,
+ ctx->gs_vtx_offset[vtx_offset_param]);
vtx_offset = LLVMBuildMul(ctx->ac.builder, gs_vtx_offset,
LLVMConstInt(ctx->i32, 4, 0), "");
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);
value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
- vtx_offset, soffset, 0, 1, 0, true, false);
+ vtx_offset, soffset, 0, ac_glc, true, false);
if (llvm_type_is_64bit(ctx, type)) {
LLVMValueRef value2;
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);
value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
ctx->i32_0, vtx_offset, soffset,
- 0, 1, 0, true, false);
+ 0, ac_glc, true, false);
return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
}
return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
offset *= 2;
offset += component;
- value[i + component] = si_llvm_load_input_gs(&ctx->abi, driver_location / 4,
+ value[i + component] = si_llvm_load_input_gs(&ctx->abi, driver_location / 4 + const_index,
vertex_index, type, offset);
}
unsigned semantic_name = info->input_semantic_name[reg->Register.Index];
if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID)
- return get_primitive_id(ctx, swizzle);
+ return si_get_primitive_id(ctx, swizzle);
if (!reg->Register.Dimension)
return NULL;
interp_fs_input(ctx, input_index, semantic_name,
semantic_index, 0, /* this param is unused */
shader->selector->info.colors_read, interp_param,
- ctx->abi.prim_mask,
+ ac_get_arg(&ctx->ac, ctx->args.prim_mask),
LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE),
&out[0]);
}
LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, SI_PARAM_ANCILLARY, 8, 4);
+ return si_unpack_param(ctx, ctx->args.ancillary, 8, 4);
}
static LLVMValueRef get_base_vertex(struct ac_shader_abi *abi)
* (for direct draws) or the CP (for indirect draws) is the
* first vertex ID, but GLSL expects 0 to be returned.
*/
- LLVMValueRef vs_state = LLVMGetParam(ctx->main_fn,
- ctx->param_vs_state_bits);
+ LLVMValueRef vs_state = ac_get_arg(&ctx->ac,
+ ctx->vs_state_bits);
LLVMValueRef indexed;
indexed = LLVMBuildLShr(ctx->ac.builder, vs_state, ctx->i32_1, "");
indexed = LLVMBuildTrunc(ctx->ac.builder, indexed, ctx->i1, "");
- return LLVMBuildSelect(ctx->ac.builder, indexed, ctx->abi.base_vertex,
+ return LLVMBuildSelect(ctx->ac.builder, indexed,
+ ac_get_arg(&ctx->ac, ctx->args.base_vertex),
ctx->i32_0, "");
}
result = ac_build_gather_values(&ctx->ac, values, 3);
} else {
- result = LLVMGetParam(ctx->main_fn, ctx->param_block_size);
+ result = ac_get_arg(&ctx->ac, ctx->block_size);
}
return result;
LLVMValueRef offset)
{
return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL,
- 0, 0, 0, true, true);
+ 0, 0, true, true);
}
static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- LLVMValueRef desc = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ LLVMValueRef desc = ac_get_arg(&ctx->ac, ctx->rw_buffers);
LLVMValueRef buf_index = LLVMConstInt(ctx->i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
LLVMValueRef resource = ac_build_load_to_sgpr(&ctx->ac, desc, buf_index);
static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- return ac_to_integer(&ctx->ac, abi->sample_coverage);
+ return ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args.sample_coverage));
}
static LLVMValueRef si_load_tess_coord(struct ac_shader_abi *abi)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
LLVMValueRef coord[4] = {
- LLVMGetParam(ctx->main_fn, ctx->param_tes_u),
- LLVMGetParam(ctx->main_fn, ctx->param_tes_v),
+ ac_get_arg(&ctx->ac, ctx->tes_u),
+ ac_get_arg(&ctx->ac, ctx->tes_v),
ctx->ac.f32_0,
ctx->ac.f32_0
};
int param = si_shader_io_get_unique_index_patch(semantic_name, 0);
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
LLVMConstInt(ctx->i32, param, 0));
}
+static LLVMValueRef load_tess_level_default(struct si_shader_context *ctx,
+ unsigned semantic_name)
+{
+ LLVMValueRef buf, slot, val[4];
+ int i, offset;
+
+ slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
+ buf = ac_get_arg(&ctx->ac, ctx->rw_buffers);
+ buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
+ offset = semantic_name == TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL ? 4 : 0;
+
+ for (i = 0; i < 4; i++)
+ val[i] = buffer_load_const(ctx, buf,
+ LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
+ return ac_build_gather_values(&ctx->ac, val, 4);
+}
+
static LLVMValueRef si_load_tess_level(struct ac_shader_abi *abi,
- unsigned varying_id)
+ unsigned varying_id,
+ bool load_default_state)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
unsigned semantic_name;
+ 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;
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
if (ctx->type == PIPE_SHADER_TESS_CTRL)
- return si_unpack_param(ctx, ctx->param_tcs_out_lds_layout, 13, 6);
+ 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
case TGSI_SEMANTIC_VERTEXID:
value = LLVMBuildAdd(ctx->ac.builder,
ctx->abi.vertex_id,
- ctx->abi.base_vertex, "");
+ ac_get_arg(&ctx->ac, ctx->args.base_vertex), "");
break;
case TGSI_SEMANTIC_VERTEXID_NOBASE:
break;
case TGSI_SEMANTIC_BASEINSTANCE:
- value = ctx->abi.start_instance;
+ value = ac_get_arg(&ctx->ac, ctx->args.start_instance);
break;
case TGSI_SEMANTIC_DRAWID:
- value = ctx->abi.draw_id;
+ value = ac_get_arg(&ctx->ac, ctx->args.draw_id);
break;
case TGSI_SEMANTIC_INVOCATIONID:
- if (ctx->type == PIPE_SHADER_TESS_CTRL)
- value = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
- else if (ctx->type == PIPE_SHADER_GEOMETRY)
- value = ctx->abi.gs_invocation_id;
- else
+ 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:
}
case TGSI_SEMANTIC_FACE:
- value = ctx->abi.front_face;
+ value = ac_get_arg(&ctx->ac, ctx->args.front_face);
break;
case TGSI_SEMANTIC_SAMPLEID:
value = load_tess_level(ctx, decl->Semantic.Name);
break;
- case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI:
- case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI:
- {
- LLVMValueRef buf, slot, val[4];
- int i, offset;
-
- slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
- buf = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
- buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
- offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;
-
- for (i = 0; i < 4; i++)
- val[i] = buffer_load_const(ctx, buf,
- LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
- value = ac_build_gather_values(&ctx->ac, val, 4);
+ 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 = get_primitive_id(ctx, 0);
+ value = si_get_primitive_id(ctx, 0);
break;
case TGSI_SEMANTIC_GRID_SIZE:
- value = ctx->abi.num_work_groups;
+ value = ac_get_arg(&ctx->ac, ctx->args.num_work_groups);
break;
case TGSI_SEMANTIC_BLOCK_SIZE:
for (int i = 0; i < 3; i++) {
values[i] = ctx->i32_0;
- if (ctx->abi.workgroup_ids[i]) {
- values[i] = ctx->abi.workgroup_ids[i];
+ 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);
}
case TGSI_SEMANTIC_THREAD_ID:
- value = ctx->abi.local_invocation_ids;
+ value = ac_get_arg(&ctx->ac, ctx->args.local_invocation_ids);
break;
case TGSI_SEMANTIC_HELPER_INVOCATION:
break;
case TGSI_SEMANTIC_SUBGROUP_SIZE:
- value = LLVMConstInt(ctx->i32, 64, 0);
+ value = LLVMConstInt(ctx->i32, ctx->ac.wave_size, 0);
break;
case TGSI_SEMANTIC_SUBGROUP_INVOCATION:
case TGSI_SEMANTIC_SUBGROUP_EQ_MASK:
{
LLVMValueRef id = ac_get_thread_id(&ctx->ac);
- id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, "");
- value = LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->i64, 1, 0), id, "");
+ 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;
}
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->i64, -2, 0);
+ value = LLVMConstInt(ctx->ac.iN_wavemask, -2, 0);
} else {
/* All bits set */
- value = LLVMConstInt(ctx->i64, -1, 0);
+ value = LLVMConstInt(ctx->ac.iN_wavemask, -1, 0);
}
- id = LLVMBuildZExt(ctx->ac.builder, id, ctx->i64, "");
+ 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;
}
- case TGSI_SEMANTIC_CS_USER_DATA:
- value = LLVMGetParam(ctx->main_fn, ctx->param_cs_user_data);
+ case TGSI_SEMANTIC_CS_USER_DATA_AMD:
+ value = ac_get_arg(&ctx->ac, ctx->cs_user_data);
break;
default:
static LLVMValueRef load_const_buffer_desc_fast_path(struct si_shader_context *ctx)
{
LLVMValueRef ptr =
- LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
+ ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
struct si_shader_selector *sel = ctx->shader->selector;
/* Do the bounds checking with a descriptor, because
desc1 = LLVMConstInt(ctx->i32,
S_008F04_BASE_ADDRESS_HI(ctx->screen->info.address32_hi), 0);
+ uint32_t rsrc3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
+
+ if (ctx->screen->info.chip_class >= GFX10)
+ rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
+ S_008F0C_OOB_SELECT(3) |
+ S_008F0C_RESOURCE_LEVEL(1);
+ else
+ rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
LLVMValueRef desc_elems[] = {
desc0,
desc1,
LLVMConstInt(ctx->i32, (sel->info.const_file_max[0] + 1) * 16, 0),
- LLVMConstInt(ctx->i32,
- S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
- S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
- S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
- S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
- S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
- S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32), 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 = LLVMGetParam(ctx->main_fn,
- ctx->param_const_and_shader_buffers);
+ 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));
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
struct si_shader_selector *sel = ctx->shader->selector;
- LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
+ LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
if (sel->info.const_buffers_declared == 1 &&
sel->info.shader_buffers_declared == 0) {
load_ssbo(struct ac_shader_abi *abi, LLVMValueRef index, bool write)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_const_and_shader_buffers);
+ LLVMValueRef rsrc_ptr = ac_get_arg(&ctx->ac,
+ ctx->const_and_shader_buffers);
index = si_llvm_bound_index(ctx, index, ctx->num_shader_buffers);
index = LLVMBuildSub(ctx->ac.builder,
buf = reg->Dimension.Index;
if (reg->Dimension.Indirect) {
- LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
+ 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,
break;
case V_028714_SPI_SHADER_32_AR:
- args->enabled_channels = 0x9; /* writemask */
- args->out[0] = values[0];
- args->out[3] = values[3];
+ if (ctx->screen->info.chip_class >= GFX10) {
+ args->enabled_channels = 0x3; /* writemask */
+ args->out[0] = values[0];
+ args->out[1] = values[3];
+ } else {
+ args->enabled_channels = 0x9; /* writemask */
+ args->out[0] = values[0];
+ args->out[3] = values[3];
+ }
break;
case V_028714_SPI_SHADER_FP16_ABGR:
unsigned chan;
unsigned const_chan;
LLVMValueRef base_elt;
- LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
SI_VS_CONST_CLIP_PLANES, 0);
LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, ptr, constbuf_index);
}
}
-static void emit_streamout_output(struct si_shader_context *ctx,
- LLVMValueRef const *so_buffers,
- LLVMValueRef const *so_write_offsets,
- struct pipe_stream_output *stream_out,
- struct si_shader_output_values *shader_out)
+void si_emit_streamout_output(struct si_shader_context *ctx,
+ LLVMValueRef const *so_buffers,
+ LLVMValueRef const *so_write_offsets,
+ struct pipe_stream_output *stream_out,
+ struct si_shader_output_values *shader_out)
{
unsigned buf_idx = stream_out->output_buffer;
unsigned start = stream_out->start_component;
vdata, num_comps,
so_write_offsets[buf_idx],
ctx->i32_0,
- stream_out->dst_offset * 4, 1, 1, true, false);
+ stream_out->dst_offset * 4, ac_glc | ac_slc);
}
/**
struct pipe_stream_output_info *so = &sel->so;
LLVMBuilderRef builder = ctx->ac.builder;
int i;
- struct lp_build_if_state if_ctx;
/* Get bits [22:16], i.e. (so_param >> 16) & 127; */
LLVMValueRef so_vtx_count =
- si_unpack_param(ctx, ctx->param_streamout_config, 16, 7);
+ si_unpack_param(ctx, ctx->streamout_config, 16, 7);
LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
/* Emit the streamout code conditionally. This actually avoids
* out-of-bounds buffer access. The hw tells us via the SGPR
* (so_vtx_count) which threads are allowed to emit streamout data. */
- lp_build_if(&if_ctx, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 6501);
{
/* The buffer offset is computed as follows:
* ByteOffset = streamout_offset[buffer_id]*4 +
*/
LLVMValueRef so_write_index =
- LLVMGetParam(ctx->main_fn,
- ctx->param_streamout_write_index);
+ ac_get_arg(&ctx->ac,
+ ctx->streamout_write_index);
/* Compute (streamout_write_index + thread_id). */
so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
* enabled buffer. */
LLVMValueRef so_write_offset[4] = {};
LLVMValueRef so_buffers[4];
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_rw_buffers);
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac,
+ ctx->rw_buffers);
for (i = 0; i < 4; i++) {
if (!so->stride[i])
so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
- LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
- ctx->param_streamout_offset[i]);
+ LLVMValueRef so_offset = ac_get_arg(&ctx->ac,
+ ctx->streamout_offset[i]);
so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
so_write_offset[i] = ac_build_imad(&ctx->ac, so_write_index,
if (stream != so->output[i].stream)
continue;
- emit_streamout_output(ctx, so_buffers, so_write_offset,
- &so->output[i], &outputs[reg]);
+ si_emit_streamout_output(ctx, so_buffers, so_write_offset,
+ &so->output[i], &outputs[reg]);
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6501);
}
static void si_export_param(struct si_shader_context *ctx, unsigned index,
return;
/* The state is in the first bit of the user SGPR. */
- LLVMValueRef cond = LLVMGetParam(ctx->main_fn, ctx->param_vs_state_bits);
+ LLVMValueRef cond = ac_get_arg(&ctx->ac, ctx->vs_state_bits);
cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->i1, "");
- struct lp_build_if_state if_ctx;
- lp_build_if(&if_ctx, &ctx->gallivm, cond);
+ ac_build_ifcc(&ctx->ac, cond, 6502);
/* Store clamped colors to alloca variables within the conditional block. */
for (unsigned i = 0; i < noutput; i++) {
addr[i][j]);
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6502);
/* Load clamped colors */
for (unsigned i = 0; i < noutput; i++) {
}
}
-/* Generate export instructions for hardware VS shader stage */
-static void si_llvm_export_vs(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
+/* Generate export instructions for hardware VS shader stage or NGG GS stage
+ * (position and parameter data only).
+ */
+void si_llvm_export_vs(struct si_shader_context *ctx,
+ struct si_shader_output_values *outputs,
+ unsigned noutput)
{
struct si_shader *shader = ctx->shader;
struct ac_export_args pos_args[4] = {};
pos_args[0].out[3] = ctx->ac.f32_1; /* W */
}
+ bool pos_writes_edgeflag = shader->selector->info.writes_edgeflag &&
+ !shader->key.as_ngg;
+
/* Write the misc vector (point size, edgeflag, layer, viewport). */
if (shader->selector->info.writes_psize ||
- shader->selector->info.writes_edgeflag ||
+ pos_writes_edgeflag ||
shader->selector->info.writes_viewport_index ||
shader->selector->info.writes_layer) {
pos_args[1].enabled_channels = shader->selector->info.writes_psize |
- (shader->selector->info.writes_edgeflag << 1) |
+ (pos_writes_edgeflag << 1) |
(shader->selector->info.writes_layer << 2);
pos_args[1].valid_mask = 0; /* EXEC mask */
if (shader->selector->info.writes_psize)
pos_args[1].out[0] = psize_value;
- if (shader->selector->info.writes_edgeflag) {
+ if (pos_writes_edgeflag) {
/* The output is a float, but the hw expects an integer
* with the first bit containing the edge flag. */
edgeflag_value = LLVMBuildFPToUI(ctx->ac.builder,
if (pos_args[i].out[0])
shader->info.nr_pos_exports++;
+ /* Navi10-14 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
+ * Setting valid_mask=1 prevents it and has no other effect.
+ */
+ if (ctx->screen->info.family == CHIP_NAVI10 ||
+ ctx->screen->info.family == CHIP_NAVI12 ||
+ ctx->screen->info.family == CHIP_NAVI14)
+ pos_args[0].valid_mask = 1;
+
pos_idx = 0;
for (i = 0; i < 4; i++) {
if (!pos_args[i].out[0])
LLVMValueRef lds_vertex_stride, lds_base;
uint64_t inputs;
- invocation_id = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
+ invocation_id = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5);
buffer = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
- buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ buffer_offset = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
lds_vertex_stride = get_tcs_in_vertex_dw_stride(ctx);
lds_base = get_tcs_in_current_patch_offset(ctx);
LLVMValueRef value = lshs_lds_load(bld_base, ctx->ac.i32, ~0, lds_ptr);
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr,
- buffer_offset, 0, 1, 0, true, false);
+ buffer_offset, 0, ac_glc);
}
}
LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
unsigned stride, outer_comps, inner_comps, i, offset;
- struct lp_build_if_state if_ctx, inner_if_ctx;
/* Add a barrier before loading tess factors from LDS. */
if (!shader->key.part.tcs.epilog.invoc0_tess_factors_are_def)
* This can't jump, because invocation 0 executes this. It should
* at least mask out the loads and stores for other invocations.
*/
- lp_build_if(&if_ctx, &ctx->gallivm,
- LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
- invocation_id, ctx->i32_0, ""));
+ ac_build_ifcc(&ctx->ac,
+ LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
+ invocation_id, ctx->i32_0, ""), 6503);
/* Determine the layout of one tess factor element in the buffer. */
switch (shader->key.part.tcs.epilog.prim_mode) {
buffer = get_tess_ring_descriptor(ctx, TCS_FACTOR_RING);
/* Get the offset. */
- tf_base = LLVMGetParam(ctx->main_fn,
- ctx->param_tcs_factor_offset);
+ tf_base = ac_get_arg(&ctx->ac,
+ ctx->tcs_factor_offset);
byteoffset = LLVMBuildMul(ctx->ac.builder, rel_patch_id,
LLVMConstInt(ctx->i32, 4 * stride, 0), "");
- lp_build_if(&inner_if_ctx, &ctx->gallivm,
- LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
- rel_patch_id, ctx->i32_0, ""));
+ ac_build_ifcc(&ctx->ac,
+ LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
+ rel_patch_id, ctx->i32_0, ""), 6504);
/* Store the dynamic HS control word. */
offset = 0;
ac_build_buffer_store_dword(&ctx->ac, buffer,
LLVMConstInt(ctx->i32, 0x80000000, 0),
1, ctx->i32_0, tf_base,
- offset, 1, 0, true, false);
+ offset, ac_glc);
offset += 4;
}
- lp_build_endif(&inner_if_ctx);
+ ac_build_endif(&ctx->ac, 6504);
/* Store the tessellation factors. */
ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
MIN2(stride, 4), byteoffset, tf_base,
- offset, 1, 0, true, false);
+ offset, ac_glc);
offset += 16;
if (vec1)
ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
stride - 4, byteoffset, tf_base,
- offset, 1, 0, true, false);
+ offset, ac_glc);
/* Store the tess factors into the offchip buffer if TES reads them. */
if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
unsigned param_outer, param_inner;
buf = get_tess_ring_descriptor(ctx, TESS_OFFCHIP_RING_TCS);
- base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
+ base = ac_get_arg(&ctx->ac, ctx->tcs_offchip_offset);
param_outer = si_shader_io_get_unique_index_patch(
TGSI_SEMANTIC_TESSOUTER, 0);
ac_build_buffer_store_dword(&ctx->ac, buf, outer_vec,
outer_comps, tf_outer_offset,
- base, 0, 1, 0, true, false);
+ base, 0, ac_glc);
if (inner_comps) {
param_inner = si_shader_io_get_unique_index_patch(
TGSI_SEMANTIC_TESSINNER, 0);
ac_build_gather_values(&ctx->ac, inner, inner_comps);
ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
inner_comps, tf_inner_offset,
- base, 0, 1, 0, true, false);
+ base, 0, ac_glc);
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6503);
}
static LLVMValueRef
si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
- unsigned param, unsigned return_index)
+ struct ac_arg param, unsigned return_index)
{
return LLVMBuildInsertValue(ctx->ac.builder, ret,
- LLVMGetParam(ctx->main_fn, param),
+ ac_get_arg(&ctx->ac, param),
return_index, "");
}
static LLVMValueRef
si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
- unsigned param, unsigned return_index)
+ struct ac_arg param, unsigned return_index)
{
LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef p = LLVMGetParam(ctx->main_fn, param);
+ LLVMValueRef p = ac_get_arg(&ctx->ac, param);
return LLVMBuildInsertValue(builder, ret,
ac_to_float(&ctx->ac, p),
static LLVMValueRef
si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
- unsigned param, unsigned return_index)
+ struct ac_arg param, unsigned return_index)
{
LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, param);
+ LLVMValueRef ptr = ac_get_arg(&ctx->ac, param);
ptr = LLVMBuildPtrToInt(builder, ptr, ctx->i32, "");
return LLVMBuildInsertValue(builder, ret, ptr, return_index, "");
}
si_copy_tcs_inputs(bld_base);
rel_patch_id = get_rel_patch_id(ctx);
- invocation_id = unpack_llvm_param(ctx, ctx->abi.tcs_rel_ids, 8, 5);
+ invocation_id = si_unpack_param(ctx, ctx->args.tcs_rel_ids, 8, 5);
tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
if (ctx->screen->info.chip_class >= GFX9) {
LLVMBasicBlockRef blocks[2] = {
LLVMGetInsertBlock(builder),
- ctx->merged_wrap_if_state.entry_block
+ ctx->merged_wrap_if_entry_block
};
LLVMValueRef values[2];
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
values[0] = rel_patch_id;
values[1] = LLVMGetUndef(ctx->i32);
unsigned vgpr;
if (ctx->screen->info.chip_class >= GFX9) {
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
8 + GFX9_SGPR_TCS_OUT_LAYOUT);
/* Tess offchip and tess factor offsets are at the beginning. */
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset, 2);
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset, 4);
vgpr = 8 + GFX9_SGPR_TCS_OUT_LAYOUT + 1;
} else {
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
GFX6_SGPR_TCS_OFFCHIP_LAYOUT);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
GFX6_SGPR_TCS_OUT_LAYOUT);
/* Tess offchip and tess factor offsets are after user SGPRs. */
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset,
GFX6_TCS_NUM_USER_SGPR);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset,
GFX6_TCS_NUM_USER_SGPR + 1);
vgpr = GFX6_TCS_NUM_USER_SGPR + 2;
}
{
LLVMValueRef ret = ctx->return_value;
- ret = si_insert_input_ptr(ctx, ret, 0, 0);
- ret = si_insert_input_ptr(ctx, ret, 1, 1);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
- ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
- ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
+ ret = si_insert_input_ptr(ctx, ret, ctx->other_const_and_shader_buffers, 0);
+ ret = si_insert_input_ptr(ctx, ret, ctx->other_samplers_and_images, 1);
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_offset, 2);
+ ret = si_insert_input_ret(ctx, ret, ctx->merged_wave_info, 3);
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_factor_offset, 4);
+ ret = si_insert_input_ret(ctx, ret, ctx->merged_scratch_offset, 5);
- ret = si_insert_input_ptr(ctx, ret, ctx->param_rw_buffers,
+ ret = si_insert_input_ptr(ctx, ret, ctx->rw_buffers,
8 + SI_SGPR_RW_BUFFERS);
ret = si_insert_input_ptr(ctx, ret,
- ctx->param_bindless_samplers_and_images,
+ ctx->bindless_samplers_and_images,
8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
- ret = si_insert_input_ret(ctx, ret, ctx->param_vs_state_bits,
+ ret = si_insert_input_ret(ctx, ret, ctx->vs_state_bits,
8 + SI_SGPR_VS_STATE_BITS);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_offchip_layout,
8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_offsets,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_offsets,
8 + GFX9_SGPR_TCS_OUT_OFFSETS);
- ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
+ ret = si_insert_input_ret(ctx, ret, ctx->tcs_out_lds_layout,
8 + GFX9_SGPR_TCS_OUT_LAYOUT);
unsigned vgpr = 8 + GFX9_TCS_NUM_USER_SGPR;
ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
- ac_to_float(&ctx->ac, ctx->abi.tcs_patch_id),
+ ac_to_float(&ctx->ac,
+ ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id)),
vgpr++, "");
ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
- ac_to_float(&ctx->ac, ctx->abi.tcs_rel_ids),
+ ac_to_float(&ctx->ac,
+ ac_get_arg(&ctx->ac, ctx->args.tcs_rel_ids)),
vgpr++, "");
ctx->return_value = ret;
}
{
LLVMValueRef ret = ctx->return_value;
- ret = si_insert_input_ptr(ctx, ret, 0, 0);
- ret = si_insert_input_ptr(ctx, ret, 1, 1);
- ret = si_insert_input_ret(ctx, ret, ctx->param_gs2vs_offset, 2);
- ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
- ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
+ ret = si_insert_input_ptr(ctx, ret, ctx->other_const_and_shader_buffers, 0);
+ ret = si_insert_input_ptr(ctx, ret, ctx->other_samplers_and_images, 1);
+ if (ctx->shader->key.as_ngg)
+ ret = si_insert_input_ptr(ctx, ret, ctx->gs_tg_info, 2);
+ else
+ ret = si_insert_input_ret(ctx, ret, ctx->gs2vs_offset, 2);
+ ret = si_insert_input_ret(ctx, ret, ctx->merged_wave_info, 3);
+ ret = si_insert_input_ret(ctx, ret, ctx->merged_scratch_offset, 5);
- ret = si_insert_input_ptr(ctx, ret, ctx->param_rw_buffers,
+ ret = si_insert_input_ptr(ctx, ret, ctx->rw_buffers,
8 + SI_SGPR_RW_BUFFERS);
ret = si_insert_input_ptr(ctx, ret,
- ctx->param_bindless_samplers_and_images,
+ ctx->bindless_samplers_and_images,
8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
+ if (ctx->screen->use_ngg) {
+ ret = si_insert_input_ptr(ctx, ret, ctx->vs_state_bits,
+ 8 + SI_SGPR_VS_STATE_BITS);
+ }
unsigned vgpr;
if (ctx->type == PIPE_SHADER_VERTEX)
else
vgpr = 8 + GFX9_TESGS_NUM_USER_SGPR;
- for (unsigned i = 0; i < 5; i++) {
- unsigned param = ctx->param_gs_vtx01_offset + i;
- ret = si_insert_input_ret_float(ctx, ret, param, vgpr++);
- }
+ ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx01_offset, vgpr++);
+ ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx23_offset, vgpr++);
+ ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_prim_id, vgpr++);
+ ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_invocation_id, vgpr++);
+ ret = si_insert_input_ret_float(ctx, ret, ctx->gs_vtx45_offset, vgpr++);
ctx->return_value = ret;
}
struct si_shader *shader = ctx->shader;
struct tgsi_shader_info *info = &shader->selector->info;
unsigned i, chan;
- LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
- ctx->param_rel_auto_id);
+ LLVMValueRef vertex_id = ac_get_arg(&ctx->ac, ctx->rel_auto_id);
LLVMValueRef vertex_dw_stride = get_tcs_in_vertex_dw_stride(ctx);
LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->ac.builder, vertex_id,
vertex_dw_stride, "");
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;
- LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
- ctx->param_es2gs_offset);
LLVMValueRef lds_base = NULL;
unsigned chan;
int i;
if (ctx->screen->info.chip_class >= GFX9 && info->num_outputs) {
unsigned itemsize_dw = es->selector->esgs_itemsize / 4;
LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
- LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
+ LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->merged_wave_info, 24, 4);
vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
LLVMBuildMul(ctx->ac.builder, wave_idx,
- LLVMConstInt(ctx->i32, 64, false), ""), "");
+ LLVMConstInt(ctx->i32, ctx->ac.wave_size, false), ""), "");
lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
LLVMConstInt(ctx->i32, itemsize_dw, 0), "");
}
ac_build_buffer_store_dword(&ctx->ac,
ctx->esgs_ring,
- out_val, 1, NULL, soffset,
+ out_val, 1, NULL,
+ ac_get_arg(&ctx->ac, ctx->es2gs_offset),
(4 * param + chan) * 4,
- 1, 1, true, true);
+ ac_glc | ac_slc | ac_swizzled);
}
}
static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
{
if (ctx->screen->info.chip_class >= GFX9)
- return si_unpack_param(ctx, ctx->param_merged_wave_info, 16, 8);
+ return si_unpack_param(ctx, ctx->merged_wave_info, 16, 8);
else
- return LLVMGetParam(ctx->main_fn, ctx->param_gs_wave_id);
+ return ac_get_arg(&ctx->ac, ctx->gs_wave_id);
}
static void emit_gs_epilogue(struct si_shader_context *ctx)
{
+ if (ctx->shader->key.as_ngg) {
+ gfx10_ngg_gs_emit_epilogue(ctx);
+ return;
+ }
+
+ if (ctx->screen->info.chip_class >= GFX10)
+ LLVMBuildFence(ctx->ac.builder, LLVMAtomicOrderingRelease, false, "");
+
ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
si_get_gs_wave_id(ctx));
if (ctx->screen->info.chip_class >= GFX9)
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
}
static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi,
}
}
- if (ctx->shader->selector->so.num_outputs)
+ if (!ctx->screen->use_ngg_streamout &&
+ ctx->shader->selector->so.num_outputs)
si_llvm_emit_streamout(ctx, outputs, i, 0);
/* Export PrimitiveID. */
if (ctx->shader->key.mono.u.vs_export_prim_id) {
outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
outputs[i].semantic_index = 0;
- outputs[i].values[0] = ac_to_float(&ctx->ac, get_primitive_id(ctx, 0));
+ outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
for (j = 1; j < 4; j++)
outputs[i].values[j] = LLVMConstReal(ctx->f32, 0);
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 waitcnt = NOOP_WAITCNT;
+ unsigned wait_flags = 0;
if (flags & TGSI_MEMBAR_THREAD_GROUP)
- waitcnt &= VM_CNT & LGKM_CNT;
+ 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))
- waitcnt &= VM_CNT;
+ wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
if (flags & TGSI_MEMBAR_SHARED)
- waitcnt &= LGKM_CNT;
+ wait_flags |= AC_WAIT_LGKM;
- if (waitcnt != NOOP_WAITCNT)
- ac_build_waitcnt(&ctx->ac, waitcnt);
+ ac_build_waitcnt(&ctx->ac, wait_flags);
}
static void clock_emit(
int input_base, input_array_size;
int chan;
int i;
- LLVMValueRef prim_mask = ctx->abi.prim_mask;
+ 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;
tmp = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
tmp = ac_build_ballot(&ctx->ac, tmp);
- tmp = LLVMBuildBitCast(builder, tmp, ctx->v2i32, "");
- emit_data->output[0] = LLVMBuildExtractElement(builder, tmp, ctx->i32_0, "");
- emit_data->output[1] = LLVMBuildExtractElement(builder, tmp, ctx->i32_1, "");
+ 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(
LLVMValueRef *addrs)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+
+ if (ctx->shader->key.as_ngg) {
+ gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
+ return;
+ }
+
struct tgsi_shader_info *info = &ctx->shader->selector->info;
struct si_shader *shader = ctx->shader;
- struct lp_build_if_state if_state;
- LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
- ctx->param_gs2vs_offset);
+ LLVMValueRef soffset = ac_get_arg(&ctx->ac, ctx->gs2vs_offset);
LLVMValueRef gs_next_vertex;
LLVMValueRef can_emit;
unsigned chan, offset;
if (use_kill) {
ac_build_kill_if_false(&ctx->ac, can_emit);
} else {
- lp_build_if(&if_state, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 6505);
}
offset = 0;
ctx->gsvs_ring[stream],
out_val, 1,
voffset, soffset, 0,
- 1, 1, true, true);
+ ac_glc | ac_slc | ac_swizzled);
}
}
}
if (!use_kill)
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6505);
}
/* Emit one vertex from the geometry shader */
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+ if (ctx->shader->key.as_ngg) {
+ LLVMBuildStore(ctx->ac.builder, ctx->ac.i32_0, ctx->gs_curprim_verts[stream]);
+ return;
+ }
+
/* Signal primitive cut */
ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
si_get_gs_wave_id(ctx));
*/
if (ctx->screen->info.chip_class == GFX6 &&
ctx->type == PIPE_SHADER_TESS_CTRL) {
- ac_build_waitcnt(&ctx->ac, LGKM_CNT & VM_CNT);
+ ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
return;
}
void si_create_function(struct si_shader_context *ctx,
const char *name,
LLVMTypeRef *returns, unsigned num_returns,
- struct si_function_info *fninfo,
unsigned max_workgroup_size)
{
- int i;
-
- si_llvm_create_func(ctx, name, returns, num_returns,
- fninfo->types, fninfo->num_params);
+ si_llvm_create_func(ctx, name, returns, num_returns);
ctx->return_value = LLVMGetUndef(ctx->return_type);
- for (i = 0; i < fninfo->num_sgpr_params; ++i) {
- LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);
-
- /* The combination of:
- * - noalias
- * - dereferenceable
- * - invariant.load
- * allows the optimization passes to move loads and reduces
- * SGPR spilling significantly.
- */
- ac_add_function_attr(ctx->ac.context, ctx->main_fn, i + 1,
- AC_FUNC_ATTR_INREG);
-
- if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
- ac_add_function_attr(ctx->ac.context, ctx->main_fn, i + 1,
- AC_FUNC_ATTR_NOALIAS);
- ac_add_attr_dereferenceable(P, UINT64_MAX);
- }
- }
-
- for (i = 0; i < fninfo->num_params; ++i) {
- if (fninfo->assign[i])
- *fninfo->assign[i] = LLVMGetParam(ctx->main_fn, i);
- }
-
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);
}
- ac_llvm_set_workgroup_size(ctx->main_fn, max_workgroup_size);
-
LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
"no-signed-zeros-fp-math",
"true");
- if (ctx->screen->debug_flags & DBG(UNSAFE_MATH)) {
- /* These were copied from some LLVM test. */
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "less-precise-fpmad",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "no-infs-fp-math",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "no-nans-fp-math",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "unsafe-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,
- struct si_function_info *fninfo)
+ struct pipe_stream_output_info *so)
{
- int i;
+ if (ctx->screen->use_ngg_streamout) {
+ if (ctx->type == PIPE_SHADER_TESS_EVAL)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ return;
+ }
/* Streamout SGPRs. */
if (so->num_outputs) {
- if (ctx->type != PIPE_SHADER_TESS_EVAL)
- ctx->param_streamout_config = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
- else
- ctx->param_streamout_config = fninfo->num_params - 1;
-
- ctx->param_streamout_write_index = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_config);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_write_index);
+ } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
}
+
/* A streamout buffer offset is loaded if the stride is non-zero. */
- for (i = 0; i < 4; i++) {
+ for (int i = 0; i < 4; i++) {
if (!so->stride[i])
continue;
- ctx->param_streamout_offset[i] = add_arg(fninfo, ARG_SGPR, ctx->ac.i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_offset[i]);
}
}
static unsigned si_get_max_workgroup_size(const struct si_shader *shader)
{
switch (shader->selector->type) {
+ case PIPE_SHADER_VERTEX:
+ case PIPE_SHADER_TESS_EVAL:
+ return shader->key.as_ngg ? 128 : 0;
+
case PIPE_SHADER_TESS_CTRL:
/* Return this so that LLVM doesn't remove s_barrier
* instructions on chips where we use s_barrier. */
- return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 64;
+ return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 0;
case PIPE_SHADER_GEOMETRY:
- return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 64;
+ return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 0;
case PIPE_SHADER_COMPUTE:
break; /* see below */
}
static void declare_const_and_shader_buffers(struct si_shader_context *ctx,
- struct si_function_info *fninfo,
bool assign_params)
{
- LLVMTypeRef const_shader_buf_type;
+ enum ac_arg_type const_shader_buf_type;
if (ctx->shader->selector->info.const_buffers_declared == 1 &&
ctx->shader->selector->info.shader_buffers_declared == 0)
- const_shader_buf_type = ctx->f32;
+ const_shader_buf_type = AC_ARG_CONST_FLOAT_PTR;
else
- const_shader_buf_type = ctx->v4i32;
-
- unsigned const_and_shader_buffers =
- add_arg(fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(const_shader_buf_type));
+ const_shader_buf_type = AC_ARG_CONST_DESC_PTR;
- if (assign_params)
- ctx->param_const_and_shader_buffers = const_and_shader_buffers;
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, const_shader_buf_type,
+ assign_params ? &ctx->const_and_shader_buffers :
+ &ctx->other_const_and_shader_buffers);
}
static void declare_samplers_and_images(struct si_shader_context *ctx,
- struct si_function_info *fninfo,
bool assign_params)
{
- unsigned samplers_and_images =
- add_arg(fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v8i32));
-
- if (assign_params)
- ctx->param_samplers_and_images = samplers_and_images;
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
+ assign_params ? &ctx->samplers_and_images :
+ &ctx->other_samplers_and_images);
}
static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
- struct si_function_info *fninfo,
bool assign_params)
{
- declare_const_and_shader_buffers(ctx, fninfo, assign_params);
- declare_samplers_and_images(ctx, fninfo, assign_params);
+ declare_const_and_shader_buffers(ctx, assign_params);
+ declare_samplers_and_images(ctx, assign_params);
}
-static void declare_global_desc_pointers(struct si_shader_context *ctx,
- struct si_function_info *fninfo)
+static void declare_global_desc_pointers(struct si_shader_context *ctx)
{
- ctx->param_rw_buffers = add_arg(fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v4i32));
- ctx->param_bindless_samplers_and_images = add_arg(fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v8i32));
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
+ &ctx->rw_buffers);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
+ &ctx->bindless_samplers_and_images);
}
-static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx,
- struct si_function_info *fninfo)
+static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx)
{
- ctx->param_vs_state_bits = add_arg(fninfo, ARG_SGPR, ctx->i32);
- add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.base_vertex);
- add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.start_instance);
- add_arg_assign(fninfo, ARG_SGPR, ctx->i32, &ctx->abi.draw_id);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ if (!ctx->shader->is_gs_copy_shader) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.base_vertex);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.start_instance);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.draw_id);
+ }
}
static void declare_vs_input_vgprs(struct si_shader_context *ctx,
- struct si_function_info *fninfo,
unsigned *num_prolog_vgprs)
{
struct si_shader *shader = ctx->shader;
- add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.vertex_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.vertex_id);
if (shader->key.as_ls) {
- ctx->param_rel_auto_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
- add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->rel_auto_id);
+ if (ctx->screen->info.chip_class >= GFX10) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ } else {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
+ }
+ } else if (ctx->screen->info.chip_class >= GFX10) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
+ &ctx->vs_prim_id); /* user vgpr or PrimID (legacy) */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
} else {
- add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
- ctx->param_vs_prim_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
}
- add_arg(fninfo, ARG_VGPR, ctx->i32); /* unused */
if (!shader->is_gs_copy_shader) {
/* Vertex load indices. */
- ctx->param_vertex_index0 = fninfo->num_params;
- for (unsigned i = 0; i < shader->selector->info.num_inputs; i++)
- add_arg(fninfo, ARG_VGPR, ctx->i32);
+ 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;
}
}
static void declare_vs_blit_inputs(struct si_shader_context *ctx,
- struct si_function_info *fninfo,
unsigned vs_blit_property)
{
- ctx->param_vs_blit_inputs = fninfo->num_params;
- add_arg(fninfo, ARG_SGPR, ctx->i32); /* i16 x1, y1 */
- add_arg(fninfo, ARG_SGPR, ctx->i32); /* i16 x2, y2 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* depth */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->vs_blit_inputs); /* i16 x1, y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* i16 x1, y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* depth */
if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* color0 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* color1 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* color2 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* color3 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color0 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color3 */
} else if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD) {
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.x1 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.y1 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.x2 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.y2 */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.z */
- add_arg(fninfo, ARG_SGPR, ctx->f32); /* texcoord.w */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.z */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.w */
}
}
-static void declare_tes_input_vgprs(struct si_shader_context *ctx,
- struct si_function_info *fninfo)
+static void declare_tes_input_vgprs(struct si_shader_context *ctx)
{
- ctx->param_tes_u = add_arg(fninfo, ARG_VGPR, ctx->f32);
- ctx->param_tes_v = add_arg(fninfo, ARG_VGPR, ctx->f32);
- ctx->param_tes_rel_patch_id = add_arg(fninfo, ARG_VGPR, ctx->i32);
- add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tes_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_u);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_v);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->tes_rel_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tes_patch_id);
}
enum {
SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY,
};
+static void add_arg_checked(struct ac_shader_args *args,
+ enum ac_arg_regfile file,
+ unsigned registers, enum ac_arg_type type,
+ struct ac_arg *arg,
+ unsigned idx)
+{
+ assert(args->arg_count == idx);
+ ac_add_arg(args, file, registers, type, arg);
+}
+
static void create_function(struct si_shader_context *ctx)
{
struct si_shader *shader = ctx->shader;
- struct si_function_info fninfo;
LLVMTypeRef returns[16+32*4];
unsigned i, num_return_sgprs;
unsigned num_returns = 0;
unsigned num_prolog_vgprs = 0;
unsigned type = ctx->type;
unsigned vs_blit_property =
- shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+ shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
/* Set MERGED shaders. */
if (ctx->screen->info.chip_class >= GFX9) {
if (shader->key.as_ls || type == PIPE_SHADER_TESS_CTRL)
type = SI_SHADER_MERGED_VERTEX_TESSCTRL; /* LS or HS */
- else if (shader->key.as_es || type == PIPE_SHADER_GEOMETRY)
+ else if (shader->key.as_es || shader->key.as_ngg || type == PIPE_SHADER_GEOMETRY)
type = SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY;
}
- LLVMTypeRef v3i32 = LLVMVectorType(ctx->i32, 3);
-
switch (type) {
case PIPE_SHADER_VERTEX:
- declare_global_desc_pointers(ctx, &fninfo);
+ declare_global_desc_pointers(ctx);
if (vs_blit_property) {
- declare_vs_blit_inputs(ctx, &fninfo, vs_blit_property);
+ declare_vs_blit_inputs(ctx, vs_blit_property);
/* VGPRs */
- declare_vs_input_vgprs(ctx, &fninfo, &num_prolog_vgprs);
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
break;
}
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
- declare_vs_specific_input_sgprs(ctx, &fninfo);
- ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v4i32));
+ 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->key.as_es) {
- ctx->param_es2gs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->es2gs_offset);
} else if (shader->key.as_ls) {
/* no extra parameters */
} else {
- if (shader->is_gs_copy_shader) {
- fninfo.num_params = ctx->param_vs_state_bits + 1;
- fninfo.num_sgpr_params = fninfo.num_params;
- }
-
/* The locations of the other parameters are assigned dynamically. */
- declare_streamout_params(ctx, &shader->selector->so,
- &fninfo);
+ declare_streamout_params(ctx, &shader->selector->so);
}
/* VGPRs */
- declare_vs_input_vgprs(ctx, &fninfo, &num_prolog_vgprs);
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
/* Return values */
if (shader->key.opt.vs_as_prim_discard_cs) {
break;
case PIPE_SHADER_TESS_CTRL: /* GFX6-GFX8 */
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_offsets = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ 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_INT, &ctx->vs_state_bits);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
/* VGPRs */
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_patch_id);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_rel_ids);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
/* param_tcs_offchip_offset and param_tcs_factor_offset are
* placed after the user SGPRs.
case SI_SHADER_MERGED_VERTEX_TESSCTRL:
/* Merged stages have 8 system SGPRs at the beginning. */
/* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
- declare_per_stage_desc_pointers(ctx, &fninfo,
+ declare_per_stage_desc_pointers(ctx,
ctx->type == PIPE_SHADER_TESS_CTRL);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_merged_scratch_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */
- add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */
-
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo,
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx,
ctx->type == PIPE_SHADER_VERTEX);
- declare_vs_specific_input_sgprs(ctx, &fninfo);
+ declare_vs_specific_input_sgprs(ctx);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_offsets = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v4i32));
+ 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);
/* VGPRs (first TCS, then VS) */
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_patch_id);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_rel_ids);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
if (ctx->type == PIPE_SHADER_VERTEX) {
- declare_vs_input_vgprs(ctx, &fninfo,
- &num_prolog_vgprs);
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
/* LS return values are inputs to the TCS main shader part. */
for (i = 0; i < 8 + GFX9_TCS_NUM_USER_SGPR; i++)
case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
/* Merged stages have 8 system SGPRs at the beginning. */
/* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
- declare_per_stage_desc_pointers(ctx, &fninfo,
+ declare_per_stage_desc_pointers(ctx,
ctx->type == PIPE_SHADER_GEOMETRY);
- ctx->param_gs2vs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_merged_scratch_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
- add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
-
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo,
- (ctx->type == PIPE_SHADER_VERTEX ||
- ctx->type == PIPE_SHADER_TESS_EVAL));
+
+ if (ctx->shader->key.as_ngg)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_tg_info);
+ else
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
+
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
+
+ declare_global_desc_pointers(ctx);
+ if (ctx->type != PIPE_SHADER_VERTEX || !vs_blit_property) {
+ declare_per_stage_desc_pointers(ctx,
+ (ctx->type == PIPE_SHADER_VERTEX ||
+ ctx->type == PIPE_SHADER_TESS_EVAL));
+ }
+
if (ctx->type == PIPE_SHADER_VERTEX) {
- declare_vs_specific_input_sgprs(ctx, &fninfo);
+ if (vs_blit_property)
+ declare_vs_blit_inputs(ctx, vs_blit_property);
+ else
+ declare_vs_specific_input_sgprs(ctx);
} else {
- ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tes_offchip_addr = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ 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->tes_offchip_addr);
/* Declare as many input SGPRs as the VS has. */
}
if (ctx->type == PIPE_SHADER_VERTEX) {
- ctx->param_vertex_buffers = add_arg(&fninfo, ARG_SGPR,
- ac_array_in_const32_addr_space(ctx->v4i32));
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
+ &ctx->vertex_buffers);
}
/* VGPRs (first GS, then VS/TES) */
- ctx->param_gs_vtx01_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
- ctx->param_gs_vtx23_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_prim_id);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_invocation_id);
- ctx->param_gs_vtx45_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx01_offset);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx23_offset);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx45_offset);
if (ctx->type == PIPE_SHADER_VERTEX) {
- declare_vs_input_vgprs(ctx, &fninfo,
- &num_prolog_vgprs);
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
} else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
- declare_tes_input_vgprs(ctx, &fninfo);
+ declare_tes_input_vgprs(ctx);
}
- if (ctx->type == PIPE_SHADER_VERTEX ||
- ctx->type == PIPE_SHADER_TESS_EVAL) {
+ if (ctx->shader->key.as_es &&
+ (ctx->type == PIPE_SHADER_VERTEX ||
+ ctx->type == PIPE_SHADER_TESS_EVAL)) {
unsigned num_user_sgprs;
if (ctx->type == PIPE_SHADER_VERTEX)
break;
case PIPE_SHADER_TESS_EVAL:
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
- ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tes_offchip_addr = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ 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->tes_offchip_addr);
if (shader->key.as_es) {
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_es2gs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->es2gs_offset);
} else {
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- declare_streamout_params(ctx, &shader->selector->so,
- &fninfo);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ declare_streamout_params(ctx, &shader->selector->so);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
}
/* VGPRs */
- declare_tes_input_vgprs(ctx, &fninfo);
+ declare_tes_input_vgprs(ctx);
break;
case PIPE_SHADER_GEOMETRY:
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
- ctx->param_gs2vs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_gs_wave_id = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_wave_id);
/* VGPRs */
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[0]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[1]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_prim_id);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[2]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[3]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[4]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->gs_vtx_offset[5]);
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.gs_invocation_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[0]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[1]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[2]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[3]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[4]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[5]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
break;
case PIPE_SHADER_FRAGMENT:
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
- add_arg_checked(&fninfo, ARG_SGPR, ctx->f32, SI_PARAM_ALPHA_REF);
- add_arg_assign_checked(&fninfo, ARG_SGPR, ctx->i32,
- &ctx->abi.prim_mask, SI_PARAM_PRIM_MASK);
-
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_SAMPLE);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_CENTER);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_PERSP_CENTROID);
- add_arg_checked(&fninfo, ARG_VGPR, v3i32, SI_PARAM_PERSP_PULL_MODEL);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_SAMPLE);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_CENTER);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->v2i32, SI_PARAM_LINEAR_CENTROID);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->f32, SI_PARAM_LINE_STIPPLE_TEX);
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
- &ctx->abi.frag_pos[0], SI_PARAM_POS_X_FLOAT);
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
- &ctx->abi.frag_pos[1], SI_PARAM_POS_Y_FLOAT);
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
- &ctx->abi.frag_pos[2], SI_PARAM_POS_Z_FLOAT);
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
- &ctx->abi.frag_pos[3], SI_PARAM_POS_W_FLOAT);
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->i32,
- &ctx->abi.front_face, SI_PARAM_FRONT_FACE);
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL,
+ SI_PARAM_ALPHA_REF);
+ add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->args.prim_mask, SI_PARAM_PRIM_MASK);
+
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_sample,
+ SI_PARAM_PERSP_SAMPLE);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
+ &ctx->args.persp_center, SI_PARAM_PERSP_CENTER);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
+ &ctx->args.persp_centroid, SI_PARAM_PERSP_CENTROID);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
+ NULL, SI_PARAM_PERSP_PULL_MODEL);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
+ &ctx->args.linear_sample, SI_PARAM_LINEAR_SAMPLE);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
+ &ctx->args.linear_center, SI_PARAM_LINEAR_CENTER);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
+ &ctx->args.linear_centroid, SI_PARAM_LINEAR_CENTROID);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_FLOAT,
+ NULL, SI_PARAM_LINE_STIPPLE_TEX);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
+ &ctx->args.frag_pos[0], SI_PARAM_POS_X_FLOAT);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
+ &ctx->args.frag_pos[1], SI_PARAM_POS_Y_FLOAT);
+ add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
+ &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);
+ 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;
- add_arg_assign_checked(&fninfo, ARG_VGPR, ctx->i32,
- &ctx->abi.ancillary, SI_PARAM_ANCILLARY);
+ 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_assign_checked(&fninfo, ARG_VGPR, ctx->f32,
- &ctx->abi.sample_coverage, SI_PARAM_SAMPLE_COVERAGE);
- add_arg_checked(&fninfo, ARG_VGPR, ctx->i32, SI_PARAM_POS_FIXED_PT);
+ 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,
+ &ctx->pos_fixed_pt, SI_PARAM_POS_FIXED_PT);
/* Color inputs from the prolog. */
if (shader->selector->info.colors_read) {
unsigned num_color_elements =
util_bitcount(shader->selector->info.colors_read);
- assert(fninfo.num_params + num_color_elements <= ARRAY_SIZE(fninfo.types));
for (i = 0; i < num_color_elements; i++)
- add_arg(&fninfo, ARG_VGPR, ctx->f32);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
num_prolog_vgprs += num_color_elements;
}
break;
case PIPE_SHADER_COMPUTE:
- declare_global_desc_pointers(ctx, &fninfo);
- declare_per_stage_desc_pointers(ctx, &fninfo, true);
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
if (shader->selector->info.uses_grid_size)
- add_arg_assign(&fninfo, ARG_SGPR, v3i32, &ctx->abi.num_work_groups);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT,
+ &ctx->args.num_work_groups);
if (shader->selector->info.uses_block_size &&
shader->selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0)
- ctx->param_block_size = add_arg(&fninfo, ARG_SGPR, v3i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT, &ctx->block_size);
unsigned cs_user_data_dwords =
- shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_DWORDS];
+ shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
if (cs_user_data_dwords) {
- ctx->param_cs_user_data = add_arg(&fninfo, ARG_SGPR,
- LLVMVectorType(ctx->i32, cs_user_data_dwords));
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, cs_user_data_dwords, AC_ARG_INT,
+ &ctx->cs_user_data);
}
+ /* Hardware SGPRs. */
for (i = 0; i < 3; i++) {
- ctx->abi.workgroup_ids[i] = NULL;
- if (shader->selector->info.uses_block_id[i])
- add_arg_assign(&fninfo, ARG_SGPR, ctx->i32, &ctx->abi.workgroup_ids[i]);
+ if (shader->selector->info.uses_block_id[i]) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->args.workgroup_ids[i]);
+ }
}
+ if (shader->selector->info.uses_subgroup_info)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.tg_size);
- add_arg_assign(&fninfo, ARG_VGPR, v3i32, &ctx->abi.local_invocation_ids);
+ /* Hardware VGPRs. */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
+ &ctx->args.local_invocation_ids);
break;
default:
assert(0 && "unimplemented shader");
return;
}
- si_create_function(ctx, "main", returns, num_returns, &fninfo,
+ si_create_function(ctx, "main", returns, num_returns,
si_get_max_workgroup_size(shader));
/* Reserve register locations for VGPR inputs the PS prolog may need. */
S_0286D0_POS_FIXED_PT_ENA(1));
}
- shader->info.num_input_sgprs = 0;
- shader->info.num_input_vgprs = 0;
-
- for (i = 0; i < fninfo.num_sgpr_params; ++i)
- shader->info.num_input_sgprs += ac_get_type_size(fninfo.types[i]) / 4;
-
- for (; i < fninfo.num_params; ++i)
- shader->info.num_input_vgprs += ac_get_type_size(fninfo.types[i]) / 4;
+ shader->info.num_input_sgprs = ctx->args.num_sgprs_used;
+ shader->info.num_input_vgprs = ctx->args.num_vgprs_used;
assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
shader->info.num_input_vgprs -= num_prolog_vgprs;
if (shader->key.as_ls || ctx->type == PIPE_SHADER_TESS_CTRL) {
- if (USE_LDS_SYMBOLS && HAVE_LLVM >= 0x0900) {
+ if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
/* The LSHS size is not known until draw time, so we append it
* at the end of whatever LDS use there may be in the rest of
* the shader (currently none, unless LLVM decides to do its
ac_declare_lds_as_pointer(&ctx->ac);
}
}
+
+ /* Unlike radv, we override these arguments in the prolog, so to the
+ * API shader they appear as normal arguments.
+ */
+ if (ctx->type == PIPE_SHADER_VERTEX) {
+ ctx->abi.vertex_id = ac_get_arg(&ctx->ac, ctx->args.vertex_id);
+ ctx->abi.instance_id = ac_get_arg(&ctx->ac, ctx->args.instance_id);
+ } else if (ctx->type == PIPE_SHADER_FRAGMENT) {
+ ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args.persp_centroid);
+ ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args.linear_centroid);
+ }
+}
+
+/* Ensure that the esgs ring is declared.
+ *
+ * We declare it with 64KB alignment as a hint that the
+ * pointer value will always be 0.
+ */
+static void declare_esgs_ring(struct si_shader_context *ctx)
+{
+ if (ctx->esgs_ring)
+ return;
+
+ assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
+
+ ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
+ ctx->ac.module, LLVMArrayType(ctx->i32, 0),
+ "esgs_ring",
+ AC_ADDR_SPACE_LDS);
+ LLVMSetLinkage(ctx->esgs_ring, LLVMExternalLinkage);
+ LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
}
/**
{
LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_rw_buffers);
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
if (ctx->shader->key.as_es || ctx->type == PIPE_SHADER_GEOMETRY) {
if (ctx->screen->info.chip_class <= GFX8) {
ctx->esgs_ring =
ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
} else {
- if (USE_LDS_SYMBOLS && HAVE_LLVM >= 0x0900) {
- /* Declare the ESGS ring as an explicit LDS symbol.
- * For monolithic shaders, we declare the ring only once.
- *
- * We declare it with 64KB alignment as a hint that the
- * pointer value will always be 0.
- */
- ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
- ctx->ac.module, LLVMArrayType(ctx->i32, 0),
- "esgs_ring",
- AC_ADDR_SPACE_LDS);
- LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
+ if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
+ /* Declare the ESGS ring as an explicit LDS symbol. */
+ declare_esgs_ring(ctx);
} else {
ac_declare_lds_as_pointer(&ctx->ac);
ctx->esgs_ring = ctx->ac.lds;
/* Limit on the stride field for <= GFX7. */
assert(stride < (1 << 14));
- num_records = 64;
+ num_records = ctx->ac.wave_size;
ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
tmp = LLVMBuildAdd(builder, tmp,
LLVMConstInt(ctx->i64,
stream_offset, 0), "");
- stream_offset += stride * 64;
+ stream_offset += stride * ctx->ac.wave_size;
ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
ring = LLVMBuildInsertElement(builder, ring,
LLVMConstInt(ctx->i32, num_records, 0),
LLVMConstInt(ctx->i32, 2, 0), "");
+
+ uint32_t rsrc3 =
+ S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+ S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
+ S_008F0C_ADD_TID_ENABLE(1);
+
+ if (ctx->ac.chip_class >= GFX10) {
+ rsrc3 |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
+ S_008F0C_OOB_SELECT(2) |
+ S_008F0C_RESOURCE_LEVEL(1);
+ } else {
+ rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
+ S_008F0C_ELEMENT_SIZE(1); /* element_size = 4 (bytes) */
+ }
+
ring = LLVMBuildInsertElement(builder, ring,
- LLVMConstInt(ctx->i32,
- S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
- S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
- S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
- S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
- S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
- S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
- S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
- S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
- S_008F0C_ADD_TID_ENABLE(1),
- 0),
+ LLVMConstInt(ctx->i32, rsrc3, false),
LLVMConstInt(ctx->i32, 3, 0), "");
ctx->gsvs_ring[stream] = ring;
static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
LLVMValueRef param_rw_buffers,
- unsigned param_pos_fixed_pt)
+ struct ac_arg param_pos_fixed_pt)
{
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef slot, desc, offset, row, bit, address[2];
#undef add_part
- struct ac_rtld_symbol lds_symbols[1];
+ struct ac_rtld_symbol lds_symbols[2];
unsigned num_lds_symbols = 0;
- if (sel && screen->info.chip_class >= GFX9 &&
- sel->type == PIPE_SHADER_GEOMETRY && !shader->is_gs_copy_shader) {
+ if (sel && screen->info.chip_class >= GFX9 && !shader->is_gs_copy_shader &&
+ (sel->type == PIPE_SHADER_GEOMETRY || shader->key.as_ngg)) {
/* We add this symbol even on LLVM <= 8 to ensure that
* shader->config.lds_size is set correctly below.
*/
sym->align = 64 * 1024;
}
+ if (shader->key.as_ngg && sel->type == PIPE_SHADER_GEOMETRY) {
+ struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
+ sym->name = "ngg_emit";
+ sym->size = shader->ngg.ngg_emit_size * 4;
+ sym->align = 4;
+ }
+
bool ok = ac_rtld_open(rtld, (struct ac_rtld_open_info){
.info = &screen->info,
+ .options = {
+ .halt_at_entry = screen->options.halt_shaders,
+ },
+ .shader_type = tgsi_processor_to_shader_stage(sel->type),
+ .wave_size = si_get_shader_wave_size(shader),
.num_parts = num_parts,
.elf_ptrs = part_elfs,
.elf_sizes = part_sizes,
{
struct ac_rtld_binary rtld;
si_shader_binary_open(screen, shader, &rtld);
- return rtld.rx_size;
+ return rtld.exec_size;
}
-
static bool si_get_external_symbol(void *data, const char *name, uint64_t *value)
{
uint64_t *scratch_va = data;
/* Enable scratch coalescing. */
*value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) |
S_008F04_SWIZZLE_ENABLE(1);
- if (HAVE_LLVM < 0x0800) {
- /* Old LLVM created an R_ABS32_HI relocation for
- * this symbol. */
- *value <<= 32;
- }
return true;
}
si_resource_reference(&shader->bo, NULL);
shader->bo = si_aligned_buffer_create(&sscreen->b,
- sscreen->cpdma_prefetch_writes_memory ?
+ sscreen->info.cpdma_prefetch_writes_memory ?
0 : SI_RESOURCE_FLAG_READ_ONLY,
PIPE_USAGE_IMMUTABLE,
align(binary.rx_size, SI_CPDMA_ALIGNMENT),
static void si_shader_dump_disassembly(struct si_screen *screen,
const struct si_shader_binary *binary,
+ enum pipe_shader_type shader_type,
+ unsigned wave_size,
struct pipe_debug_callback *debug,
const char *name, FILE *file)
{
if (!ac_rtld_open(&rtld_binary, (struct ac_rtld_open_info){
.info = &screen->info,
+ .shader_type = tgsi_processor_to_shader_stage(shader_type),
+ .wave_size = wave_size,
.num_parts = 1,
.elf_ptrs = &binary->elf_buffer,
.elf_sizes = &binary->elf_size }))
if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm, &nbytes))
goto out;
- fprintf(file, "Shader %s disassembly:\n", name);
- if (nbytes > INT_MAX) {
- fprintf(file, "too long\n");
+ if (nbytes > INT_MAX)
goto out;
- }
-
- fprintf(file, "%*s", (int)nbytes, disasm);
if (debug && debug->debug_message) {
/* Very long debug messages are cut off, so send the
"Shader Disassembly End");
}
+ if (file) {
+ fprintf(file, "Shader %s disassembly:\n", name);
+ fprintf(file, "%*s", (int)nbytes, disasm);
+ }
+
out:
ac_rtld_close(&rtld_binary);
}
unsigned lds_per_wave = 0;
unsigned max_simd_waves;
- max_simd_waves = ac_get_max_simd_waves(sscreen->info.family);
+ max_simd_waves = sscreen->info.max_wave64_per_simd;
/* Compute LDS usage for PS. */
switch (shader->selector->type) {
unsigned max_workgroup_size =
si_get_max_workgroup_size(shader);
lds_per_wave = (conf->lds_size * lds_increment) /
- DIV_ROUND_UP(max_workgroup_size, 64);
+ DIV_ROUND_UP(max_workgroup_size,
+ sscreen->compute_wave_size);
}
break;
+ default:;
}
/* Compute the per-SIMD wave counts. */
if (conf->num_sgprs) {
max_simd_waves =
MIN2(max_simd_waves,
- ac_get_num_physical_sgprs(sscreen->info.chip_class) / conf->num_sgprs);
+ sscreen->info.num_physical_sgprs_per_simd / conf->num_sgprs);
}
- if (conf->num_vgprs)
- max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
+ if (conf->num_vgprs) {
+ /* Always print wave limits as Wave64, so that we can compare
+ * Wave32 and Wave64 with shader-db fairly. */
+ unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd;
+ max_simd_waves = MIN2(max_simd_waves, max_vgprs / conf->num_vgprs);
+ }
- /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
- * 16KB makes some SIMDs unoccupied). */
+ /* LDS is 64KB per CU (4 SIMDs) on GFX6-9, which is 16KB per SIMD (usage above
+ * 16KB makes some SIMDs unoccupied).
+ *
+ * LDS is 128KB in WGP mode and 64KB in CU mode. Assume the WGP mode is used.
+ */
+ unsigned max_lds_size = sscreen->info.chip_class >= GFX10 ? 128*1024 : 64*1024;
+ unsigned max_lds_per_simd = max_lds_size / 4;
if (lds_per_wave)
- max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
+ max_simd_waves = MIN2(max_simd_waves, max_lds_per_simd / lds_per_wave);
shader->info.max_simd_waves = max_simd_waves;
}
{
const struct ac_shader_config *conf = &shader->config;
+ if (screen->options.debug_disassembly)
+ si_shader_dump_disassembly(screen, &shader->binary,
+ shader->selector->type,
+ si_get_shader_wave_size(shader),
+ debug, "main", NULL);
+
pipe_debug_message(debug, SHADER_INFO,
"Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
"LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
static void si_shader_dump_stats(struct si_screen *sscreen,
struct si_shader *shader,
- unsigned processor,
FILE *file,
bool check_debug_option)
{
const struct ac_shader_config *conf = &shader->config;
if (!check_debug_option ||
- si_can_dump_shader(sscreen, processor)) {
- if (processor == PIPE_SHADER_FRAGMENT) {
+ si_can_dump_shader(sscreen, shader->selector->type)) {
+ if (shader->selector->type == PIPE_SHADER_FRAGMENT) {
fprintf(file, "*** SHADER CONFIG ***\n"
"SPI_PS_INPUT_ADDR = 0x%04x\n"
"SPI_PS_INPUT_ENA = 0x%04x\n",
}
}
-const char *si_get_shader_name(const struct si_shader *shader, unsigned processor)
+const char *si_get_shader_name(const struct si_shader *shader)
{
- switch (processor) {
+ switch (shader->selector->type) {
case PIPE_SHADER_VERTEX:
if (shader->key.as_es)
return "Vertex Shader as ES";
return "Vertex Shader as LS";
else if (shader->key.opt.vs_as_prim_discard_cs)
return "Vertex Shader as Primitive Discard CS";
+ else if (shader->key.as_ngg)
+ return "Vertex Shader as ESGS";
else
return "Vertex Shader as VS";
case PIPE_SHADER_TESS_CTRL:
case PIPE_SHADER_TESS_EVAL:
if (shader->key.as_es)
return "Tessellation Evaluation Shader as ES";
+ else if (shader->key.as_ngg)
+ return "Tessellation Evaluation Shader as ESGS";
else
return "Tessellation Evaluation Shader as VS";
case PIPE_SHADER_GEOMETRY:
}
void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
- struct pipe_debug_callback *debug, unsigned processor,
+ struct pipe_debug_callback *debug,
FILE *file, bool check_debug_option)
{
+ enum pipe_shader_type shader_type = shader->selector->type;
+
if (!check_debug_option ||
- si_can_dump_shader(sscreen, processor))
- si_dump_shader_key(processor, shader, file);
+ si_can_dump_shader(sscreen, shader_type))
+ si_dump_shader_key(shader, file);
if (!check_debug_option && shader->binary.llvm_ir_string) {
if (shader->previous_stage &&
shader->previous_stage->binary.llvm_ir_string) {
fprintf(file, "\n%s - previous stage - LLVM IR:\n\n",
- si_get_shader_name(shader, processor));
+ si_get_shader_name(shader));
fprintf(file, "%s\n", shader->previous_stage->binary.llvm_ir_string);
}
fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
- si_get_shader_name(shader, processor));
+ si_get_shader_name(shader));
fprintf(file, "%s\n", shader->binary.llvm_ir_string);
}
if (!check_debug_option ||
- (si_can_dump_shader(sscreen, processor) &&
+ (si_can_dump_shader(sscreen, shader_type) &&
!(sscreen->debug_flags & DBG(NO_ASM)))) {
- fprintf(file, "\n%s:\n", si_get_shader_name(shader, processor));
+ unsigned wave_size = si_get_shader_wave_size(shader);
+
+ fprintf(file, "\n%s:\n", si_get_shader_name(shader));
if (shader->prolog)
si_shader_dump_disassembly(sscreen, &shader->prolog->binary,
- debug, "prolog", file);
+ shader_type, wave_size, debug, "prolog", file);
if (shader->previous_stage)
si_shader_dump_disassembly(sscreen, &shader->previous_stage->binary,
- debug, "previous stage", file);
+ shader_type, wave_size, debug, "previous stage", file);
if (shader->prolog2)
si_shader_dump_disassembly(sscreen, &shader->prolog2->binary,
- debug, "prolog2", file);
+ shader_type, wave_size, debug, "prolog2", file);
- si_shader_dump_disassembly(sscreen, &shader->binary, debug, "main", file);
+ si_shader_dump_disassembly(sscreen, &shader->binary, shader_type,
+ wave_size, debug, "main", file);
if (shader->epilog)
si_shader_dump_disassembly(sscreen, &shader->epilog->binary,
- debug, "epilog", file);
+ shader_type, wave_size, debug, "epilog", file);
fprintf(file, "\n");
}
- si_shader_dump_stats(sscreen, shader, processor, file,
- check_debug_option);
+ si_shader_dump_stats(sscreen, shader, file, check_debug_option);
}
static int si_compile_llvm(struct si_screen *sscreen,
struct ac_llvm_compiler *compiler,
LLVMModuleRef mod,
struct pipe_debug_callback *debug,
- unsigned processor,
+ enum pipe_shader_type shader_type,
+ unsigned wave_size,
const char *name,
bool less_optimized)
{
unsigned count = p_atomic_inc_return(&sscreen->num_compilations);
- if (si_can_dump_shader(sscreen, processor)) {
+ if (si_can_dump_shader(sscreen, shader_type)) {
fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
if (!(sscreen->debug_flags & (DBG(NO_IR) | DBG(PREOPT_IR)))) {
if (!si_replace_shader(count, binary)) {
unsigned r = si_llvm_compile(mod, binary, compiler, debug,
- less_optimized);
+ less_optimized, wave_size);
if (r)
return r;
}
struct ac_rtld_binary rtld;
if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
.info = &sscreen->info,
+ .shader_type = tgsi_processor_to_shader_stage(shader_type),
+ .wave_size = wave_size,
.num_parts = 1,
.elf_ptrs = &binary->elf_buffer,
.elf_sizes = &binary->elf_size }))
shader->selector = gs_selector;
shader->is_gs_copy_shader = true;
- si_init_shader_ctx(&ctx, sscreen, compiler);
+ si_init_shader_ctx(&ctx, sscreen, compiler,
+ si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false),
+ false);
ctx.shader = shader;
ctx.type = PIPE_SHADER_VERTEX;
/* Fetch the vertex stream ID.*/
LLVMValueRef stream_id;
- if (gs_selector->so.num_outputs)
- stream_id = si_unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
+ if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs)
+ stream_id = si_unpack_param(&ctx, ctx.streamout_config, 24, 2);
else
stream_id = ctx.i32_0;
ac_build_buffer_load(&ctx.ac,
ctx.gsvs_ring[0], 1,
ctx.i32_0, voffset,
- soffset, 0, 1, 1,
+ soffset, 0, ac_glc | ac_slc,
true, false);
}
}
/* Streamout and exports. */
- if (gs_selector->so.num_outputs) {
+ if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs) {
si_llvm_emit_streamout(&ctx, outputs,
gsinfo->num_outputs,
stream);
if (si_compile_llvm(sscreen, &ctx.shader->binary,
&ctx.shader->config, ctx.compiler,
ctx.ac.module,
- debug, PIPE_SHADER_GEOMETRY,
+ debug, PIPE_SHADER_GEOMETRY, ctx.ac.wave_size,
"GS Copy Shader", false) == 0) {
if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY))
fprintf(stderr, "GS Copy Shader:\n");
- si_shader_dump(sscreen, ctx.shader, debug,
- PIPE_SHADER_GEOMETRY, stderr, true);
+ si_shader_dump(sscreen, ctx.shader, debug, stderr, true);
if (!ctx.shader->config.scratch_bytes_per_wave)
ok = si_shader_binary_upload(sscreen, ctx.shader, 0);
fprintf(f, "}\n");
}
-static void si_dump_shader_key(unsigned processor, const struct si_shader *shader,
- FILE *f)
+static void si_dump_shader_key(const struct si_shader *shader, FILE *f)
{
const struct si_shader_key *key = &shader->key;
+ enum pipe_shader_type shader_type = shader->selector->type;
fprintf(f, "SHADER KEY\n");
- switch (processor) {
+ switch (shader_type) {
case PIPE_SHADER_VERTEX:
si_dump_shader_key_vs(key, &key->part.vs.prolog,
"part.vs.prolog", f);
fprintf(f, " as_es = %u\n", key->as_es);
fprintf(f, " as_ls = %u\n", key->as_ls);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
fprintf(f, " mono.u.vs_export_prim_id = %u\n",
key->mono.u.vs_export_prim_id);
fprintf(f, " opt.vs_as_prim_discard_cs = %u\n",
case PIPE_SHADER_TESS_EVAL:
fprintf(f, " as_es = %u\n", key->as_es);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
fprintf(f, " mono.u.vs_export_prim_id = %u\n",
key->mono.u.vs_export_prim_id);
break;
"part.gs.vs_prolog", f);
}
fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
+ fprintf(f, " part.gs.prolog.gfx9_prev_is_vs = %u\n", key->part.gs.prolog.gfx9_prev_is_vs);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
break;
case PIPE_SHADER_COMPUTE:
fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
+ fprintf(f, " part.ps.prolog.samplemask_log_ps_iter = %u\n", key->part.ps.prolog.samplemask_log_ps_iter);
fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
+ fprintf(f, " mono.u.ps.interpolate_at_sample_force_center = %u\n", key->mono.u.ps.interpolate_at_sample_force_center);
+ fprintf(f, " mono.u.ps.fbfetch_msaa = %u\n", key->mono.u.ps.fbfetch_msaa);
+ fprintf(f, " mono.u.ps.fbfetch_is_1D = %u\n", key->mono.u.ps.fbfetch_is_1D);
+ fprintf(f, " mono.u.ps.fbfetch_layered = %u\n", key->mono.u.ps.fbfetch_layered);
break;
default:
assert(0);
}
- if ((processor == PIPE_SHADER_GEOMETRY ||
- processor == PIPE_SHADER_TESS_EVAL ||
- processor == PIPE_SHADER_VERTEX) &&
+ if ((shader_type == PIPE_SHADER_GEOMETRY ||
+ shader_type == PIPE_SHADER_TESS_EVAL ||
+ shader_type == PIPE_SHADER_VERTEX) &&
!key->as_es && !key->as_ls) {
fprintf(f, " opt.kill_outputs = 0x%"PRIx64"\n", key->opt.kill_outputs);
fprintf(f, " opt.clip_disable = %u\n", key->opt.clip_disable);
static void si_init_shader_ctx(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler)
+ struct ac_llvm_compiler *compiler,
+ unsigned wave_size,
+ bool nir)
{
struct lp_build_tgsi_context *bld_base;
- si_llvm_context_init(ctx, sscreen, compiler);
+ 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;
}
static void si_init_exec_from_input(struct si_shader_context *ctx,
- unsigned param, unsigned bitoffset)
+ struct ac_arg param, unsigned bitoffset)
{
LLVMValueRef args[] = {
- LLVMGetParam(ctx->main_fn, param),
+ ac_get_arg(&ctx->ac, param),
LLVMConstInt(ctx->i32, bitoffset, 0),
};
ac_build_intrinsic(&ctx->ac,
return sel->vs_needs_prolog || key->ls_vgpr_fix;
}
-static bool si_compile_tgsi_main(struct si_shader_context *ctx)
+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 bool si_compile_tgsi_main(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;
ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
else if (shader->key.opt.vs_as_prim_discard_cs)
ctx->abi.emit_outputs = si_llvm_emit_prim_discard_cs_epilogue;
+ else if (shader->key.as_ngg)
+ 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;
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.load_patch_vertices_in = si_load_patch_vertices_in;
if (shader->key.as_es)
ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
+ else if (shader->key.as_ngg)
+ 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->load_input = declare_input_fs;
ctx->abi.emit_outputs = si_llvm_return_fs_outputs;
bld_base->emit_epilogue = si_tgsi_emit_epilogue;
- ctx->abi.lookup_interp_param = si_nir_lookup_interp_param;
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;
ctx->abi.emit_kill = si_llvm_emit_kill;
break;
case PIPE_SHADER_COMPUTE:
create_function(ctx);
preload_ring_buffers(ctx);
+ if (ctx->type == PIPE_SHADER_TESS_CTRL &&
+ sel->tcs_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, "");
+ }
+ }
+
+ if (ctx->type == PIPE_SHADER_GEOMETRY) {
+ for (unsigned i = 0; i < 4; i++) {
+ ctx->gs_next_vertex[i] =
+ ac_build_alloca(&ctx->ac, ctx->i32, "");
+ }
+ if (shader->key.as_ngg) {
+ for (unsigned i = 0; i < 4; ++i) {
+ ctx->gs_curprim_verts[i] =
+ ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
+ ctx->gs_generated_prims[i] =
+ ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
+ }
+
+ unsigned scratch_size = 8;
+ if (sel->so.num_outputs)
+ scratch_size = 44;
+
+ LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, scratch_size);
+ ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
+ ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+ LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(ai32));
+ LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
+
+ ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx->ac.module,
+ LLVMArrayType(ctx->i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
+ LLVMSetLinkage(ctx->gs_ngg_emit, LLVMExternalLinkage);
+ LLVMSetAlignment(ctx->gs_ngg_emit, 4);
+ }
+ }
+
+ if (ctx->type != PIPE_SHADER_GEOMETRY &&
+ (shader->key.as_ngg && !shader->key.as_es)) {
+ /* Unconditionally declare scratch space base for streamout and
+ * vertex compaction. Whether space is actually allocated is
+ * determined during linking / PM4 creation.
+ *
+ * Add an extra dword per vertex to ensure an odd stride, which
+ * avoids bank conflicts for SoA accesses.
+ */
+ 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);
+ }
+
/* For GFX9 merged shaders:
* - Set EXEC for the first shader. If the prolog is present, set
* EXEC there instead.
*
* For monolithic merged shaders, the first shader is wrapped in an
* if-block together with its prolog in si_build_wrapper_function.
+ *
+ * NGG vertex and tess eval shaders running as the last
+ * vertex/geometry stage handle execution explicitly using
+ * if-statements.
*/
if (ctx->screen->info.chip_class >= GFX9) {
if (!shader->is_monolithic &&
(ctx->type == PIPE_SHADER_VERTEX &&
!si_vs_needs_prolog(sel, &shader->key.part.vs.prolog)))) {
si_init_exec_from_input(ctx,
- ctx->param_merged_wave_info, 0);
+ ctx->merged_wave_info, 0);
} else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
- ctx->type == PIPE_SHADER_GEOMETRY) {
- if (!shader->is_monolithic)
+ ctx->type == PIPE_SHADER_GEOMETRY ||
+ (shader->key.as_ngg && !shader->key.as_es)) {
+ LLVMValueRef thread_enabled;
+ bool nested_barrier;
+
+ if (!shader->is_monolithic ||
+ (ctx->type == PIPE_SHADER_TESS_EVAL &&
+ (shader->key.as_ngg && !shader->key.as_es)))
ac_init_exec_full_mask(&ctx->ac);
- LLVMValueRef num_threads = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
- LLVMValueRef ena =
- LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), num_threads, "");
- lp_build_if(&ctx->merged_wrap_if_state, &ctx->gallivm, ena);
+ if (ctx->type == PIPE_SHADER_TESS_CTRL ||
+ ctx->type == PIPE_SHADER_GEOMETRY) {
+ if (ctx->type == PIPE_SHADER_GEOMETRY && shader->key.as_ngg) {
+ gfx10_ngg_gs_emit_prologue(ctx);
+ nested_barrier = false;
+ } else {
+ nested_barrier = true;
+ }
- /* The barrier must execute for all shaders in a
- * threadgroup.
- *
- * Execute the barrier inside the conditional block,
- * so that empty waves can jump directly to s_endpgm,
- * which will also signal the barrier.
- *
- * If the shader is TCS and the TCS epilog is present
- * and contains a barrier, it will wait there and then
- * reach s_endpgm.
- */
- si_llvm_emit_barrier(NULL, bld_base, NULL);
- }
- }
+ thread_enabled = si_is_gs_thread(ctx);
+ } else {
+ thread_enabled = si_is_es_thread(ctx);
+ nested_barrier = false;
+ }
- if (ctx->type == PIPE_SHADER_TESS_CTRL &&
- sel->tcs_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, "");
- }
- }
+ ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
+ ctx->merged_wrap_if_label = 11500;
+ ac_build_ifcc(&ctx->ac, thread_enabled, ctx->merged_wrap_if_label);
- if (ctx->type == PIPE_SHADER_GEOMETRY) {
- int i;
- for (i = 0; i < 4; i++) {
- ctx->gs_next_vertex[i] =
- ac_build_alloca(&ctx->ac, ctx->i32, "");
+ if (nested_barrier) {
+ /* Execute a barrier before the second shader in
+ * a merged shader.
+ *
+ * Execute the barrier inside the conditional block,
+ * so that empty waves can jump directly to s_endpgm,
+ * which will also signal the barrier.
+ *
+ * This is possible in gfx9, because an empty wave
+ * for the second shader does not participate in
+ * the epilogue. With NGG, empty waves may still
+ * be required to export data (e.g. GS output vertices),
+ * so we cannot let them exit early.
+ *
+ * If the shader is TCS and the TCS epilog is present
+ * and contains a barrier, it will wait there and then
+ * reach s_endpgm.
+ */
+ si_llvm_emit_barrier(NULL, bld_base, NULL);
+ }
}
}
return false;
}
} else {
- if (!si_nir_build_llvm(ctx, sel->nir)) {
+ 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;
}
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;
if (shader_out->selector->type == PIPE_SHADER_TESS_CTRL) {
key->vs_prolog.as_ls = 1;
} else if (shader_out->selector->type == PIPE_SHADER_GEOMETRY) {
key->vs_prolog.as_es = 1;
key->vs_prolog.num_merged_next_stage_vgprs = 5;
+ } else if (shader_out->key.as_ngg) {
+ key->vs_prolog.num_merged_next_stage_vgprs = 5;
}
/* Enable loading the InstanceID VGPR. */
/* BCOLORs are stored after the last input. */
key->ps_prolog.num_interp_inputs = info->num_inputs;
key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
- shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
+ if (separate_prolog)
+ shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
}
for (unsigned i = 0; i < 2; i++) {
switch (location) {
case TGSI_INTERPOLATE_LOC_SAMPLE:
key->ps_prolog.color_interp_vgpr_index[i] = 0;
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_SAMPLE_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_PERSP_SAMPLE_ENA(1);
+ }
break;
case TGSI_INTERPOLATE_LOC_CENTER:
key->ps_prolog.color_interp_vgpr_index[i] = 2;
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_CENTER_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_PERSP_CENTER_ENA(1);
+ }
break;
case TGSI_INTERPOLATE_LOC_CENTROID:
key->ps_prolog.color_interp_vgpr_index[i] = 4;
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_CENTROID_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_PERSP_CENTROID_ENA(1);
+ }
break;
default:
assert(0);
case TGSI_INTERPOLATE_LOC_SAMPLE:
key->ps_prolog.color_interp_vgpr_index[i] =
separate_prolog ? 6 : 9;
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_SAMPLE_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_LINEAR_SAMPLE_ENA(1);
+ }
break;
case TGSI_INTERPOLATE_LOC_CENTER:
key->ps_prolog.color_interp_vgpr_index[i] =
separate_prolog ? 8 : 11;
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_CENTER_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_LINEAR_CENTER_ENA(1);
+ }
break;
case TGSI_INTERPOLATE_LOC_CENTROID:
key->ps_prolog.color_interp_vgpr_index[i] =
separate_prolog ? 10 : 13;
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_CENTROID_ENA(1);
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |=
+ S_0286CC_LINEAR_CENTROID_ENA(1);
+ }
break;
default:
assert(0);
union si_shader_part_key *key)
{
unsigned num_sgprs, num_vgprs;
- struct si_function_info fninfo;
LLVMBuilderRef builder = ctx->ac.builder;
LLVMTypeRef returns[48];
LLVMValueRef func, ret;
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
if (ctx->screen->info.chip_class >= GFX9) {
if (key->gs_prolog.states.gfx9_prev_is_vs)
}
for (unsigned i = 0; i < num_sgprs; ++i) {
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
returns[i] = ctx->i32;
}
for (unsigned i = 0; i < num_vgprs; ++i) {
- add_arg(&fninfo, ARG_VGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
returns[num_sgprs + i] = ctx->f32;
}
/* Create the function. */
si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
- &fninfo, 0);
+ 0);
func = ctx->main_fn;
/* Set the full EXEC mask for the prolog, because we are only fiddling
if (key->gs_prolog.states.tri_strip_adj_fix) {
/* Remap the input vertices for every other primitive. */
- const unsigned gfx6_vtx_params[6] = {
- num_sgprs,
- num_sgprs + 1,
- num_sgprs + 3,
- num_sgprs + 4,
- num_sgprs + 5,
- num_sgprs + 6
+ const struct ac_arg gfx6_vtx_params[6] = {
+ { .used = true, .arg_index = num_sgprs },
+ { .used = true, .arg_index = num_sgprs + 1 },
+ { .used = true, .arg_index = num_sgprs + 3 },
+ { .used = true, .arg_index = num_sgprs + 4 },
+ { .used = true, .arg_index = num_sgprs + 5 },
+ { .used = true, .arg_index = num_sgprs + 6 },
};
- const unsigned gfx9_vtx_params[3] = {
- num_sgprs,
- num_sgprs + 1,
- num_sgprs + 4,
+ const struct ac_arg gfx9_vtx_params[3] = {
+ { .used = true, .arg_index = num_sgprs },
+ { .used = true, .arg_index = num_sgprs + 1 },
+ { .used = true, .arg_index = num_sgprs + 4 },
};
LLVMValueRef vtx_in[6], vtx_out[6];
LLVMValueRef prim_id, rotate;
}
} else {
for (unsigned i = 0; i < 6; i++)
- vtx_in[i] = LLVMGetParam(func, gfx6_vtx_params[i]);
+ vtx_in[i] = ac_get_arg(&ctx->ac, gfx6_vtx_params[i]);
}
prim_id = LLVMGetParam(func, num_sgprs + 2);
out = LLVMBuildOr(builder, vtx_out[i*2], hi, "");
out = ac_to_float(&ctx->ac, out);
ret = LLVMBuildInsertValue(builder, ret, out,
- gfx9_vtx_params[i], "");
+ gfx9_vtx_params[i].arg_index, "");
}
} else {
for (unsigned i = 0; i < 6; i++) {
out = ac_to_float(&ctx->ac, vtx_out[i]);
ret = LLVMBuildInsertValue(builder, ret, out,
- gfx6_vtx_params[i], "");
+ gfx6_vtx_params[i].arg_index, "");
}
}
}
/* PS epilog has one arg per color component; gfx9 merged shader
* prologs need to forward 32 user SGPRs.
*/
- struct si_function_info fninfo;
LLVMValueRef initial[64], out[64];
LLVMTypeRef function_type;
unsigned num_first_params;
unsigned num_out, initial_num_out;
- MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
- MAYBE_UNUSED unsigned initial_num_out_sgpr; /* used in debug checks */
+ ASSERTED unsigned num_out_sgpr; /* used in debug checks */
+ ASSERTED unsigned initial_num_out_sgpr; /* used in debug checks */
unsigned num_sgprs, num_vgprs;
unsigned gprs;
- struct lp_build_if_state if_state;
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
for (unsigned i = 0; i < num_parts; ++i) {
ac_add_function_attr(ctx->ac.context, parts[i], -1,
gprs = 0;
while (gprs < num_sgprs + num_vgprs) {
- LLVMValueRef param = LLVMGetParam(parts[main_part], fninfo.num_params);
+ LLVMValueRef param = LLVMGetParam(parts[main_part], ctx->args.arg_count);
LLVMTypeRef type = LLVMTypeOf(param);
unsigned size = ac_get_type_size(type) / 4;
- add_arg(&fninfo, gprs < num_sgprs ? ARG_SGPR : ARG_VGPR, type);
+ /* This is going to get casted anyways, so we don't have to
+ * have the exact same type. But we do have to preserve the
+ * pointer-ness so that LLVM knows about it.
+ */
+ enum ac_arg_type arg_type = AC_ARG_INT;
+ if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
+ arg_type = AC_ARG_CONST_PTR;
+ }
+
+ ac_add_arg(&ctx->args, gprs < num_sgprs ? AC_ARG_SGPR : AC_ARG_VGPR,
+ size, arg_type, NULL);
assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
assert(gprs + size <= num_sgprs + num_vgprs &&
unreachable("unexpected type");
}
- si_create_function(ctx, "wrapper", returns, num_returns, &fninfo,
+ si_create_function(ctx, "wrapper", returns, num_returns,
si_get_max_workgroup_size(ctx->shader));
if (is_merged_shader(ctx))
num_out = 0;
num_out_sgpr = 0;
- for (unsigned i = 0; i < fninfo.num_params; ++i) {
+ for (unsigned i = 0; i < ctx->args.arg_count; ++i) {
LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
LLVMTypeRef param_type = LLVMTypeOf(param);
- LLVMTypeRef out_type = i < fninfo.num_sgpr_params ? ctx->i32 : ctx->f32;
+ LLVMTypeRef out_type = ctx->args.args[i].file == AC_ARG_SGPR ? ctx->i32 : ctx->f32;
unsigned size = ac_get_type_size(param_type) / 4;
if (size == 1) {
builder, param, LLVMConstInt(ctx->i32, j, 0), "");
}
- if (i < fninfo.num_sgpr_params)
+ if (ctx->args.args[i].file == AC_ARG_SGPR)
num_out_sgpr = num_out;
}
initial_num_out_sgpr = num_out_sgpr;
/* Now chain the parts. */
- LLVMValueRef ret;
+ LLVMValueRef ret = NULL;
for (unsigned part = 0; part < num_parts; ++part) {
LLVMValueRef in[48];
LLVMTypeRef ret_type;
/* Merged shaders are executed conditionally depending
* on the number of enabled threads passed in the input SGPRs. */
- if (is_merged_shader(ctx) && part == 0) {
+ if (is_multi_part_shader(ctx) && part == 0) {
LLVMValueRef ena, count = initial[3];
count = LLVMBuildAnd(builder, count,
LLVMConstInt(ctx->i32, 0x7f, 0), "");
ena = LLVMBuildICmp(builder, LLVMIntULT,
ac_get_thread_id(&ctx->ac), count, "");
- lp_build_if(&if_state, &ctx->gallivm, ena);
+ ac_build_ifcc(&ctx->ac, ena, 6506);
}
/* Derive arguments for the next part from outputs of the
out_idx += param_size;
}
- ret = LLVMBuildCall(builder, parts[part], in, num_params, "");
+ ret = ac_build_call(&ctx->ac, parts[part], in, num_params);
- if (is_merged_shader(ctx) &&
+ if (is_multi_part_shader(ctx) &&
part + 1 == next_shader_first_part) {
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6506);
/* The second half of the merged shader should use
* the inputs from the toplevel (wrapper) function,
sel->info.num_memory_instructions > 1000;
}
+static struct nir_shader *get_nir_shader(struct si_shader_selector *sel,
+ bool *free_nir)
+{
+ *free_nir = false;
+
+ if (sel->nir) {
+ return sel->nir;
+ } else if (sel->nir_binary) {
+ struct pipe_screen *screen = &sel->screen->b;
+ const void *options =
+ screen->get_compiler_options(screen, PIPE_SHADER_IR_NIR,
+ sel->type);
+
+ struct blob_reader blob_reader;
+ blob_reader_init(&blob_reader, sel->nir_binary, sel->nir_size);
+ *free_nir = true;
+ return nir_deserialize(NULL, options, &blob_reader);
+ }
+ return NULL;
+}
+
int si_compile_tgsi_shader(struct si_screen *sscreen,
struct ac_llvm_compiler *compiler,
struct si_shader *shader,
{
struct si_shader_selector *sel = shader->selector;
struct si_shader_context ctx;
+ bool free_nir;
+ struct nir_shader *nir = get_nir_shader(sel, &free_nir);
int r = -1;
/* Dump TGSI code before doing TGSI->LLVM conversion in case the
* conversion fails. */
- if (si_can_dump_shader(sscreen, sel->info.processor) &&
+ 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(sel->nir, stderr);
+ nir_print_shader(nir, stderr);
si_dump_streamout(&sel->so);
}
- si_init_shader_ctx(&ctx, sscreen, compiler);
- si_llvm_context_set_tgsi(&ctx, shader);
+ si_init_shader_ctx(&ctx, sscreen, compiler, si_get_shader_wave_size(shader),
+ nir != NULL);
+ si_llvm_context_set_ir(&ctx, shader, nir);
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)) {
+ if (!si_compile_tgsi_main(&ctx, nir, free_nir)) {
si_llvm_dispose(&ctx);
return -1;
}
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;
}
parts[3] = ctx.main_fn;
/* VS as LS main part */
+ nir = get_nir_shader(ls, &free_nir);
struct si_shader shader_ls = {};
shader_ls.selector = ls;
shader_ls.key.as_ls = 1;
shader_ls.key.mono = shader->key.mono;
shader_ls.key.opt = shader->key.opt;
shader_ls.is_monolithic = true;
- si_llvm_context_set_tgsi(&ctx, &shader_ls);
+ si_llvm_context_set_ir(&ctx, &shader_ls, nir);
- if (!si_compile_tgsi_main(&ctx)) {
+ if (!si_compile_tgsi_main(&ctx, nir, free_nir)) {
si_llvm_dispose(&ctx);
return -1;
}
memset(&gs_prolog_key, 0, sizeof(gs_prolog_key));
gs_prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
gs_prolog_key.gs_prolog.is_monolithic = true;
+ gs_prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
si_build_gs_prolog_function(&ctx, &gs_prolog_key);
gs_prolog = ctx.main_fn;
/* ES main part */
+ nir = get_nir_shader(es, &free_nir);
struct si_shader shader_es = {};
shader_es.selector = es;
shader_es.key.as_es = 1;
+ shader_es.key.as_ngg = shader->key.as_ngg;
shader_es.key.mono = shader->key.mono;
shader_es.key.opt = shader->key.opt;
shader_es.is_monolithic = true;
- si_llvm_context_set_tgsi(&ctx, &shader_es);
+ si_llvm_context_set_ir(&ctx, &shader_es, nir);
- if (!si_compile_tgsi_main(&ctx)) {
+ if (!si_compile_tgsi_main(&ctx, nir, free_nir)) {
si_llvm_dispose(&ctx);
return -1;
}
/* Compile to bytecode. */
r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler,
- ctx.ac.module, debug, ctx.type,
- si_get_shader_name(shader, ctx.type),
+ ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
+ si_get_shader_name(shader),
si_should_optimize_less(compiler, shader->selector));
si_llvm_dispose(&ctx);
if (r) {
* LLVM 3.9svn has this bug.
*/
if (sel->type == PIPE_SHADER_COMPUTE) {
- unsigned wave_size = 64;
- unsigned max_vgprs = 256;
- unsigned max_sgprs = sscreen->info.chip_class >= GFX8 ? 800 : 512;
+ unsigned wave_size = sscreen->compute_wave_size;
+ unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd *
+ (wave_size == 32 ? 2 : 1);
+ unsigned max_sgprs = sscreen->info.num_physical_sgprs_per_simd;
unsigned max_sgprs_per_wave = 128;
- unsigned max_block_threads = si_get_max_workgroup_size(shader);
- unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
- unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
+ unsigned simds_per_tg = 4; /* assuming WGP mode on gfx10 */
+ unsigned threads_per_tg = si_get_max_workgroup_size(shader);
+ unsigned waves_per_tg = DIV_ROUND_UP(threads_per_tg, wave_size);
+ unsigned waves_per_simd = DIV_ROUND_UP(waves_per_tg, simds_per_tg);
- max_vgprs = max_vgprs / min_waves_per_simd;
- max_sgprs = MIN2(max_sgprs / min_waves_per_simd, max_sgprs_per_wave);
+ max_vgprs = max_vgprs / waves_per_simd;
+ max_sgprs = MIN2(max_sgprs / waves_per_simd, max_sgprs_per_wave);
if (shader->config.num_sgprs > max_sgprs ||
shader->config.num_vgprs > max_vgprs) {
/* Calculate the number of fragment input VGPRs. */
if (ctx.type == PIPE_SHADER_FRAGMENT) {
- shader->info.num_input_vgprs = 0;
- shader->info.face_vgpr_index = -1;
- shader->info.ancillary_vgpr_index = -1;
-
- if (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_PERSP_PULL_MODEL_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 3;
- if (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 2;
- if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_POS_X_FLOAT_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_POS_Y_FLOAT_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_POS_Z_FLOAT_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_FRONT_FACE_ENA(shader->config.spi_ps_input_addr)) {
- shader->info.face_vgpr_index = shader->info.num_input_vgprs;
- shader->info.num_input_vgprs += 1;
- }
- if (G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr)) {
- shader->info.ancillary_vgpr_index = shader->info.num_input_vgprs;
- shader->info.num_input_vgprs += 1;
- }
- if (G_0286CC_SAMPLE_COVERAGE_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
- if (G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr))
- shader->info.num_input_vgprs += 1;
+ shader->info.num_input_vgprs = ac_get_fs_input_vgpr_cnt(&shader->config,
+ &shader->info.face_vgpr_index,
+ &shader->info.ancillary_vgpr_index);
}
si_calculate_max_simd_waves(shader);
{
struct si_shader_part *result;
- mtx_lock(&sscreen->shader_parts_mutex);
+ simple_mtx_lock(&sscreen->shader_parts_mutex);
/* Find existing. */
for (result = *list; result; result = result->next) {
if (memcmp(&result->key, key, sizeof(*key)) == 0) {
- mtx_unlock(&sscreen->shader_parts_mutex);
+ simple_mtx_unlock(&sscreen->shader_parts_mutex);
return result;
}
}
result->key = *key;
struct si_shader shader = {};
- struct si_shader_context ctx;
-
- si_init_shader_ctx(&ctx, sscreen, compiler);
- ctx.shader = &shader;
- ctx.type = type;
switch (type) {
case PIPE_SHADER_VERTEX:
shader.key.as_ls = key->vs_prolog.as_ls;
shader.key.as_es = key->vs_prolog.as_es;
+ shader.key.as_ngg = key->vs_prolog.as_ngg;
break;
case PIPE_SHADER_TESS_CTRL:
assert(!prolog);
break;
case PIPE_SHADER_GEOMETRY:
assert(prolog);
+ shader.key.as_ngg = key->gs_prolog.as_ngg;
break;
case PIPE_SHADER_FRAGMENT:
if (prolog)
unreachable("bad shader part");
}
+ 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);
+ ctx.shader = &shader;
+ ctx.type = type;
+
build(&ctx, key);
/* Compile. */
si_llvm_optimize_module(&ctx);
if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler,
- ctx.ac.module, debug, ctx.type, name, false)) {
+ ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
+ name, false)) {
FREE(result);
result = NULL;
goto out;
out:
si_llvm_dispose(&ctx);
- mtx_unlock(&sscreen->shader_parts_mutex);
+ simple_mtx_unlock(&sscreen->shader_parts_mutex);
return result;
}
static void si_build_vs_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key)
{
- struct si_function_info fninfo;
LLVMTypeRef *returns;
LLVMValueRef ret, func;
int num_returns, i;
unsigned first_vs_vgpr = key->vs_prolog.num_merged_next_stage_vgprs;
unsigned num_input_vgprs = key->vs_prolog.num_merged_next_stage_vgprs + 4;
+ struct ac_arg input_sgpr_param[key->vs_prolog.num_input_sgprs];
+ struct ac_arg input_vgpr_param[9];
LLVMValueRef input_vgprs[9];
unsigned num_all_input_regs = key->vs_prolog.num_input_sgprs +
num_input_vgprs;
unsigned user_sgpr_base = key->vs_prolog.num_merged_next_stage_vgprs ? 8 : 0;
- si_init_function_info(&fninfo);
+ 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;
/* Declare input and output SGPRs. */
for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &input_sgpr_param[i]);
returns[num_returns++] = ctx->i32;
}
+ struct ac_arg merged_wave_info = input_sgpr_param[3];
+
/* Preloaded VGPRs (outputs must be floats) */
for (i = 0; i < num_input_vgprs; i++) {
- add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &input_vgprs[i]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &input_vgpr_param[i]);
returns[num_returns++] = ctx->f32;
}
/* 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, &fninfo, 0);
+ si_create_function(ctx, "vs_prolog", returns, num_returns, 0);
func = ctx->main_fn;
+ for (i = 0; i < num_input_vgprs; i++) {
+ input_vgprs[i] = ac_get_arg(&ctx->ac, input_vgpr_param[i]);
+ }
+
if (key->vs_prolog.num_merged_next_stage_vgprs) {
if (!key->vs_prolog.is_monolithic)
- si_init_exec_from_input(ctx, 3, 0);
+ si_init_exec_from_input(ctx, merged_wave_info, 0);
if (key->vs_prolog.as_ls &&
- ctx->screen->has_ls_vgpr_init_bug) {
+ ctx->screen->info.has_ls_vgpr_init_bug) {
/* If there are no HS threads, SPI loads the LS VGPRs
* starting at VGPR 0. Shift them back to where they
* belong.
*/
LLVMValueRef has_hs_threads =
LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
- si_unpack_param(ctx, 3, 8, 8),
+ si_unpack_param(ctx, input_sgpr_param[3], 8, 8),
ctx->i32_0, "");
for (i = 4; i > 0; --i) {
}
unsigned vertex_id_vgpr = first_vs_vgpr;
- unsigned instance_id_vgpr = first_vs_vgpr + (key->vs_prolog.as_ls ? 2 : 1);
+ unsigned instance_id_vgpr =
+ ctx->screen->info.chip_class >= GFX10 ?
+ first_vs_vgpr + 3 :
+ first_vs_vgpr + (key->vs_prolog.as_ls ? 2 : 1);
ctx->abi.vertex_id = input_vgprs[vertex_id_vgpr];
ctx->abi.instance_id = input_vgprs[instance_id_vgpr];
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 =
index = ac_to_float(&ctx->ac, index);
ret = LLVMBuildInsertValue(ctx->ac.builder, ret, index,
- fninfo.num_params + i, "");
+ ctx->args.arg_count + i, "");
}
si_llvm_build_ret(ctx, ret);
union si_shader_part_key *key)
{
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct si_function_info fninfo;
- LLVMValueRef func;
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
if (ctx->screen->info.chip_class >= GFX9) {
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32); /* wave info */
- ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* wave info */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->tcs_factor_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ 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, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->tcs_out_lds_layout);
} else {
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
- }
-
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* VGPR gap */
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* VGPR gap */
- unsigned tess_factors_idx =
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* patch index within the wave (REL_PATCH_ID) */
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* invocation ID within the patch */
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* LDS offset where tess factors should be loaded from */
-
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ 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, NULL);
+ 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_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->tcs_factor_offset);
+ }
+
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* VGPR gap */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* VGPR gap */
+ struct ac_arg rel_patch_id; /* patch index within the wave (REL_PATCH_ID) */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &rel_patch_id);
+ struct ac_arg invocation_id; /* invocation ID within the patch */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &invocation_id);
+ struct ac_arg tcs_out_current_patch_data_offset; /* LDS offset where tess factors should be loaded from */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
+ &tcs_out_current_patch_data_offset);
+
+ struct ac_arg tess_factors[6];
for (unsigned i = 0; i < 6; i++)
- add_arg(&fninfo, ARG_VGPR, ctx->i32); /* tess factors */
+ 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, &fninfo,
- ctx->screen->info.chip_class >= GFX7 ? 128 : 64);
+ si_create_function(ctx, "tcs_epilog", NULL, 0,
+ ctx->screen->info.chip_class >= GFX7 ? 128 : 0);
ac_declare_lds_as_pointer(&ctx->ac);
- func = ctx->main_fn;
LLVMValueRef invoc0_tess_factors[6];
for (unsigned i = 0; i < 6; i++)
- invoc0_tess_factors[i] = LLVMGetParam(func, tess_factors_idx + 3 + i);
+ invoc0_tess_factors[i] = ac_get_arg(&ctx->ac, tess_factors[i]);
si_write_tess_factors(bld_base,
- LLVMGetParam(func, tess_factors_idx),
- LLVMGetParam(func, tess_factors_idx + 1),
- LLVMGetParam(func, tess_factors_idx + 2),
+ 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),
invoc0_tess_factors, invoc0_tess_factors + 4);
LLVMBuildRetVoid(ctx->ac.builder);
struct pipe_debug_callback *debug)
{
if (sscreen->info.chip_class >= GFX9) {
- struct si_shader *es_main_part =
- shader->key.part.gs.es->main_shader_part_es;
+ struct si_shader *es_main_part;
+ enum pipe_shader_type es_type = shader->key.part.gs.es->type;
+
+ if (shader->key.as_ngg)
+ es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
+ else
+ es_main_part = shader->key.part.gs.es->main_shader_part_es;
- if (shader->key.part.gs.es->type == PIPE_SHADER_VERTEX &&
+ if (es_type == PIPE_SHADER_VERTEX &&
!si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
&shader->key.part.gs.vs_prolog))
return false;
union si_shader_part_key prolog_key;
memset(&prolog_key, 0, sizeof(prolog_key));
prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+ prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
shader->prolog2 = si_get_shader_part(sscreen, &sscreen->gs_prologs,
PIPE_SHADER_GEOMETRY, true,
static void si_build_ps_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key)
{
- struct si_function_info fninfo;
LLVMValueRef ret, func;
int num_returns, i, num_color_channels;
assert(si_need_ps_prolog(key));
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
/* Declare inputs. */
- for (i = 0; i < key->ps_prolog.num_input_sgprs; i++)
- add_arg(&fninfo, ARG_SGPR, ctx->i32);
-
- for (i = 0; i < key->ps_prolog.num_input_vgprs; i++)
- add_arg(&fninfo, ARG_VGPR, ctx->f32);
+ LLVMTypeRef return_types[AC_MAX_ARGS];
+ num_returns = 0;
+ num_color_channels = util_bitcount(key->ps_prolog.colors_read);
+ assert(key->ps_prolog.num_input_sgprs +
+ key->ps_prolog.num_input_vgprs +
+ num_color_channels <= AC_MAX_ARGS);
+ for (i = 0; i < key->ps_prolog.num_input_sgprs; i++) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ return_types[num_returns++] = ctx->i32;
+
+ }
+
+ struct ac_arg pos_fixed_pt;
+ struct ac_arg ancillary;
+ struct ac_arg param_sample_mask;
+ for (i = 0; i < key->ps_prolog.num_input_vgprs; i++) {
+ struct ac_arg *arg = NULL;
+ if (i == key->ps_prolog.ancillary_vgpr_index) {
+ arg = &ancillary;
+ } else if (i == key->ps_prolog.ancillary_vgpr_index + 1) {
+ arg = ¶m_sample_mask;
+ } else if (i == key->ps_prolog.num_input_vgprs - 1) {
+ /* POS_FIXED_PT is always last. */
+ arg = &pos_fixed_pt;
+ }
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, arg);
+ return_types[num_returns++] = ctx->f32;
+ }
/* Declare outputs (same as inputs + add colors if needed) */
- num_returns = fninfo.num_params;
- num_color_channels = util_bitcount(key->ps_prolog.colors_read);
for (i = 0; i < num_color_channels; i++)
- fninfo.types[num_returns++] = ctx->f32;
+ return_types[num_returns++] = ctx->f32;
/* Create the function. */
- si_create_function(ctx, "ps_prolog", fninfo.types, num_returns,
- &fninfo, 0);
+ si_create_function(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,
* but it will prevent the compiler from overwriting them unintentionally.
*/
ret = ctx->return_value;
- for (i = 0; i < fninfo.num_params; i++) {
+ for (i = 0; i < ctx->args.arg_count; i++) {
LLVMValueRef p = LLVMGetParam(func, i);
ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
}
/* Polygon stippling. */
if (key->ps_prolog.states.poly_stipple) {
- /* POS_FIXED_PT is always last. */
- unsigned pos = key->ps_prolog.num_input_sgprs +
- key->ps_prolog.num_input_vgprs - 1;
LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
- si_llvm_emit_polygon_stipple(ctx, list, pos);
+ si_llvm_emit_polygon_stipple(ctx, list, pos_fixed_pt);
}
if (key->ps_prolog.states.bc_optimize_for_persp ||
while (writemask) {
unsigned chan = u_bit_scan(&writemask);
ret = LLVMBuildInsertValue(ctx->ac.builder, ret, color[chan],
- fninfo.num_params + color_out_idx++, "");
+ ctx->args.arg_count + color_out_idx++, "");
}
}
assert(key->ps_prolog.states.samplemask_log_ps_iter < ARRAY_SIZE(ps_iter_masks));
uint32_t ps_iter_mask = ps_iter_masks[key->ps_prolog.states.samplemask_log_ps_iter];
- unsigned ancillary_vgpr = key->ps_prolog.num_input_sgprs +
- key->ps_prolog.ancillary_vgpr_index;
- LLVMValueRef sampleid = si_unpack_param(ctx, ancillary_vgpr, 8, 4);
- LLVMValueRef samplemask = LLVMGetParam(func, ancillary_vgpr + 1);
+ LLVMValueRef sampleid = si_unpack_param(ctx, ancillary, 8, 4);
+ LLVMValueRef samplemask = ac_get_arg(&ctx->ac, param_sample_mask);
samplemask = ac_to_integer(&ctx->ac, samplemask);
samplemask = LLVMBuildAnd(
samplemask = ac_to_float(&ctx->ac, samplemask);
ret = LLVMBuildInsertValue(ctx->ac.builder, ret, samplemask,
- ancillary_vgpr + 1, "");
+ param_sample_mask.arg_index, "");
}
/* Tell LLVM to insert WQM instruction sequence when needed. */
union si_shader_part_key *key)
{
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct si_function_info fninfo;
LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
int i;
struct si_ps_exports exp = {};
- si_init_function_info(&fninfo);
+ memset(&ctx->args, 0, sizeof(ctx->args));
/* Declare input SGPRs. */
- ctx->param_rw_buffers = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- ctx->param_bindless_samplers_and_images = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- ctx->param_const_and_shader_buffers = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- ctx->param_samplers_and_images = add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
- add_arg_checked(&fninfo, ARG_SGPR, ctx->f32, SI_PARAM_ALPHA_REF);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->rw_buffers);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->bindless_samplers_and_images);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->const_and_shader_buffers);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ &ctx->samplers_and_images);
+ add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT,
+ NULL, SI_PARAM_ALPHA_REF);
/* Declare input VGPRs. */
unsigned required_num_params =
- fninfo.num_sgpr_params +
+ ctx->args.num_sgprs_used +
util_bitcount(key->ps_epilog.colors_written) * 4 +
key->ps_epilog.writes_z +
key->ps_epilog.writes_stencil +
key->ps_epilog.writes_samplemask;
required_num_params = MAX2(required_num_params,
- fninfo.num_sgpr_params + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
+ ctx->args.num_sgprs_used + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
- while (fninfo.num_params < required_num_params)
- add_arg(&fninfo, ARG_VGPR, ctx->f32);
+ while (ctx->args.arg_count < required_num_params)
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
/* Create the function. */
- si_create_function(ctx, "ps_epilog", NULL, 0, &fninfo, 0);
+ si_create_function(ctx, "ps_epilog", NULL, 0, 0);
/* Disable elimination of unused inputs. */
ac_llvm_add_target_dep_function_attr(ctx->main_fn,
"InitialPSInputAddr", 0xffffff);
/* Process colors. */
- unsigned vgpr = fninfo.num_sgpr_params;
+ unsigned vgpr = ctx->args.num_sgprs_used;
unsigned colors_written = key->ps_epilog.colors_written;
int last_color_export = -1;
color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
si_export_mrt_color(bld_base, color, mrt,
- fninfo.num_params - 1,
+ ctx->args.arg_count - 1,
mrt == last_color_export, &exp);
}
shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
if (shader->selector->type == PIPE_SHADER_COMPUTE &&
- si_get_max_workgroup_size(shader) > 64) {
+ si_get_max_workgroup_size(shader) > sscreen->compute_wave_size) {
si_multiwave_lds_size_workaround(sscreen,
&shader->config.lds_size);
}
/* LS, ES, VS are compiled on demand if the main part hasn't been
* compiled for that stage.
*
+ * GS are compiled on demand if the main part hasn't been compiled
+ * for the chosen NGG-ness.
+ *
* Vertex shaders are compiled on demand when a vertex fetch
* workaround must be applied.
*/
shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
shader->info.num_input_vgprs);
break;
+ default:;
}
/* Update SGPR and VGPR counts. */
si_calculate_max_simd_waves(shader);
}
- if (sscreen->info.chip_class >= GFX9 && sel->type == PIPE_SHADER_GEOMETRY)
+ if (shader->key.as_ngg) {
+ assert(!shader->key.as_es && !shader->key.as_ls);
+ gfx10_ngg_calculate_subgroup_info(shader);
+ } else if (sscreen->info.chip_class >= GFX9 && sel->type == PIPE_SHADER_GEOMETRY) {
gfx9_get_gs_info(shader->previous_stage_sel, sel, &shader->gs_info);
+ }
si_fix_resource_usage(sscreen, shader);
- si_shader_dump(sscreen, shader, debug, sel->info.processor,
- stderr, true);
+ si_shader_dump(sscreen, shader, debug, stderr, true);
/* Upload. */
if (!si_shader_binary_upload(sscreen, shader, 0)) {
projects / mesa.git / blobdiff