#include <llvm-c/Core.h>
#include <llvm-c/TargetMachine.h>
#include <llvm-c/Transforms/Scalar.h>
-#if HAVE_LLVM >= 0x0700
#include <llvm-c/Transforms/Utils.h>
-#endif
#include "sid.h"
-#include "gfx9d.h"
#include "ac_binary.h"
#include "ac_llvm_util.h"
#include "ac_llvm_build.h"
gl_shader_stage stage;
LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS * 4];
+ uint64_t float16_shaded_mask;
uint64_t input_mask;
uint64_t output_mask;
/* Make sure that the data fits in LDS. This assumes the shaders only
* use LDS for the inputs and outputs.
*/
- hardware_lds_size = ctx->options->chip_class >= CIK ? 65536 : 32768;
+ hardware_lds_size = 32768;
+
+ /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
+ * threadgroup, even though there is more than 32 KiB LDS.
+ *
+ * Test: dEQP-VK.tessellation.shader_input_output.barrier
+ */
+ if (ctx->options->chip_class >= GFX7 && ctx->options->family != CHIP_STONEY)
+ hardware_lds_size = 65536;
+
num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
/* Make sure the output data fits in the offchip buffer */
num_patches = MIN2(num_patches, (ctx->options->tess_offchip_block_dw_size * 4) / output_patch_size);
*/
num_patches = MIN2(num_patches, 40);
- /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
- if (ctx->options->chip_class == SI) {
+ /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
+ if (ctx->options->chip_class == GFX6) {
unsigned one_wave = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp);
num_patches = MIN2(num_patches, one_wave);
}
struct arg_info {
LLVMTypeRef types[MAX_ARGS];
LLVMValueRef *assign[MAX_ARGS];
- unsigned array_params_mask;
uint8_t count;
uint8_t sgpr_count;
uint8_t num_sgprs_used;
}
}
-static inline void
-add_array_arg(struct arg_info *info, LLVMTypeRef type, LLVMValueRef *param_ptr)
-{
- info->array_params_mask |= (1 << info->count);
- add_arg(info, ARG_SGPR, type, param_ptr);
-}
-
static void assign_arguments(LLVMValueRef main_function,
struct arg_info *info)
{
LLVMSetFunctionCallConv(main_function, RADEON_LLVM_AMDGPU_CS);
for (unsigned i = 0; i < args->sgpr_count; ++i) {
+ LLVMValueRef P = LLVMGetParam(main_function, i);
+
ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_INREG);
- if (args->array_params_mask & (1 << i)) {
- LLVMValueRef P = LLVMGetParam(main_function, i);
+ if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
ac_add_attr_dereferenceable(P, UINT64_MAX);
}
options->address32_hi);
}
- if (max_workgroup_size) {
- ac_llvm_add_target_dep_function_attr(main_function,
- "amdgpu-max-work-group-size",
- max_workgroup_size);
- }
+ ac_llvm_set_workgroup_size(main_function, max_workgroup_size);
+
if (options->unsafe_math) {
/* These were copied from some LLVM test. */
LLVMAddTargetDependentFunctionAttr(main_function,
static void
set_loc(struct radv_userdata_info *ud_info, uint8_t *sgpr_idx,
- uint8_t num_sgprs, bool indirect)
+ uint8_t num_sgprs)
{
ud_info->sgpr_idx = *sgpr_idx;
ud_info->num_sgprs = num_sgprs;
- ud_info->indirect = indirect;
*sgpr_idx += num_sgprs;
}
&ctx->shader_info->user_sgprs_locs.shader_data[idx];
assert(ud_info);
- set_loc(ud_info, sgpr_idx, num_sgprs, false);
+ set_loc(ud_info, sgpr_idx, num_sgprs);
}
static void
set_loc_shader_ptr(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx)
{
- bool use_32bit_pointers = HAVE_32BIT_POINTERS &&
- idx != AC_UD_SCRATCH_RING_OFFSETS;
+ bool use_32bit_pointers = idx != AC_UD_SCRATCH_RING_OFFSETS;
set_loc_shader(ctx, idx, sgpr_idx, use_32bit_pointers ? 1 : 2);
}
static void
-set_loc_desc(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx,
- bool indirect)
+set_loc_desc(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx)
{
struct radv_userdata_locations *locs =
&ctx->shader_info->user_sgprs_locs;
struct radv_userdata_info *ud_info = &locs->descriptor_sets[idx];
assert(ud_info);
- set_loc(ud_info, sgpr_idx, HAVE_32BIT_POINTERS ? 1 : 2, indirect);
+ set_loc(ud_info, sgpr_idx, 1);
- if (!indirect)
- locs->descriptor_sets_enabled |= 1 << idx;
+ locs->descriptor_sets_enabled |= 1 << idx;
}
struct user_sgpr_info {
bool need_ring_offsets;
bool indirect_all_descriptor_sets;
+ uint8_t remaining_sgprs;
};
static bool needs_view_index_sgpr(struct radv_shader_context *ctx,
uint8_t count = 0;
if (ctx->shader_info->info.vs.has_vertex_buffers)
- count += HAVE_32BIT_POINTERS ? 1 : 2;
+ count++;
count += ctx->shader_info->info.vs.needs_draw_id ? 3 : 2;
return count;
}
+static void allocate_inline_push_consts(struct radv_shader_context *ctx,
+ struct user_sgpr_info *user_sgpr_info)
+{
+ uint8_t remaining_sgprs = user_sgpr_info->remaining_sgprs;
+
+ /* Only supported if shaders use push constants. */
+ if (ctx->shader_info->info.min_push_constant_used == UINT8_MAX)
+ return;
+
+ /* Only supported if shaders don't have indirect push constants. */
+ if (ctx->shader_info->info.has_indirect_push_constants)
+ return;
+
+ /* Only supported for 32-bit push constants. */
+ if (!ctx->shader_info->info.has_only_32bit_push_constants)
+ return;
+
+ uint8_t num_push_consts =
+ (ctx->shader_info->info.max_push_constant_used -
+ ctx->shader_info->info.min_push_constant_used) / 4;
+
+ /* Check if the number of user SGPRs is large enough. */
+ if (num_push_consts < remaining_sgprs) {
+ ctx->shader_info->info.num_inline_push_consts = num_push_consts;
+ } else {
+ ctx->shader_info->info.num_inline_push_consts = remaining_sgprs;
+ }
+
+ /* Clamp to the maximum number of allowed inlined push constants. */
+ if (ctx->shader_info->info.num_inline_push_consts > AC_MAX_INLINE_PUSH_CONSTS)
+ ctx->shader_info->info.num_inline_push_consts = AC_MAX_INLINE_PUSH_CONSTS;
+
+ if (ctx->shader_info->info.num_inline_push_consts == num_push_consts &&
+ !ctx->shader_info->info.loads_dynamic_offsets) {
+ /* Disable the default push constants path if all constants are
+ * inlined and if shaders don't use dynamic descriptors.
+ */
+ ctx->shader_info->info.loads_push_constants = false;
+ }
+
+ ctx->shader_info->info.base_inline_push_consts =
+ ctx->shader_info->info.min_push_constant_used / 4;
+}
+
static void allocate_user_sgprs(struct radv_shader_context *ctx,
gl_shader_stage stage,
bool has_previous_stage,
user_sgpr_count++;
if (ctx->shader_info->info.loads_push_constants)
- user_sgpr_count += HAVE_32BIT_POINTERS ? 1 : 2;
+ user_sgpr_count++;
+
+ if (ctx->streamout_buffers)
+ user_sgpr_count++;
uint32_t available_sgprs = ctx->options->chip_class >= GFX9 && stage != MESA_SHADER_COMPUTE ? 32 : 16;
uint32_t remaining_sgprs = available_sgprs - user_sgpr_count;
uint32_t num_desc_set =
util_bitcount(ctx->shader_info->info.desc_set_used_mask);
- if (remaining_sgprs / (HAVE_32BIT_POINTERS ? 1 : 2) < num_desc_set) {
+ if (remaining_sgprs < num_desc_set) {
user_sgpr_info->indirect_all_descriptor_sets = true;
+ user_sgpr_info->remaining_sgprs = remaining_sgprs - 1;
+ } else {
+ user_sgpr_info->remaining_sgprs = remaining_sgprs - num_desc_set;
}
+
+ allocate_inline_push_consts(ctx, user_sgpr_info);
}
static void
declare_global_input_sgprs(struct radv_shader_context *ctx,
- gl_shader_stage stage,
- bool has_previous_stage,
- gl_shader_stage previous_stage,
const struct user_sgpr_info *user_sgpr_info,
struct arg_info *args,
LLVMValueRef *desc_sets)
{
LLVMTypeRef type = ac_array_in_const32_addr_space(ctx->ac.i8);
- unsigned num_sets = ctx->options->layout ?
- ctx->options->layout->num_sets : 0;
- unsigned stage_mask = 1 << stage;
-
- if (has_previous_stage)
- stage_mask |= 1 << previous_stage;
/* 1 for each descriptor set */
if (!user_sgpr_info->indirect_all_descriptor_sets) {
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- add_array_arg(args, type,
- &ctx->descriptor_sets[i]);
- }
+ uint32_t mask = ctx->shader_info->info.desc_set_used_mask;
+
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ add_arg(args, ARG_SGPR, type, &ctx->descriptor_sets[i]);
}
} else {
- add_array_arg(args, ac_array_in_const32_addr_space(type), desc_sets);
+ add_arg(args, ARG_SGPR, ac_array_in_const32_addr_space(type),
+ desc_sets);
}
if (ctx->shader_info->info.loads_push_constants) {
/* 1 for push constants and dynamic descriptors */
- add_array_arg(args, type, &ctx->abi.push_constants);
+ add_arg(args, ARG_SGPR, type, &ctx->abi.push_constants);
}
+ for (unsigned i = 0; i < ctx->shader_info->info.num_inline_push_consts; i++) {
+ add_arg(args, ARG_SGPR, ctx->ac.i32,
+ &ctx->abi.inline_push_consts[i]);
+ }
+ ctx->abi.num_inline_push_consts = ctx->shader_info->info.num_inline_push_consts;
+ ctx->abi.base_inline_push_consts = ctx->shader_info->info.base_inline_push_consts;
+
if (ctx->shader_info->info.so.num_outputs) {
add_arg(args, ARG_SGPR,
ac_array_in_const32_addr_space(ctx->ac.v4i32),
}
static void
-set_global_input_locs(struct radv_shader_context *ctx, gl_shader_stage stage,
- bool has_previous_stage, gl_shader_stage previous_stage,
+set_global_input_locs(struct radv_shader_context *ctx,
const struct user_sgpr_info *user_sgpr_info,
LLVMValueRef desc_sets, uint8_t *user_sgpr_idx)
{
- unsigned num_sets = ctx->options->layout ?
- ctx->options->layout->num_sets : 0;
- unsigned stage_mask = 1 << stage;
-
- if (has_previous_stage)
- stage_mask |= 1 << previous_stage;
+ uint32_t mask = ctx->shader_info->info.desc_set_used_mask;
if (!user_sgpr_info->indirect_all_descriptor_sets) {
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- set_loc_desc(ctx, i, user_sgpr_idx, false);
- } else
- ctx->descriptor_sets[i] = NULL;
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ set_loc_desc(ctx, i, user_sgpr_idx);
}
} else {
set_loc_shader_ptr(ctx, AC_UD_INDIRECT_DESCRIPTOR_SETS,
user_sgpr_idx);
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- ctx->descriptor_sets[i] =
- ac_build_load_to_sgpr(&ctx->ac,
- desc_sets,
- LLVMConstInt(ctx->ac.i32, i, false));
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ ctx->descriptor_sets[i] =
+ ac_build_load_to_sgpr(&ctx->ac, desc_sets,
+ LLVMConstInt(ctx->ac.i32, i, false));
- } else
- ctx->descriptor_sets[i] = NULL;
}
+
ctx->shader_info->need_indirect_descriptor_sets = true;
}
set_loc_shader_ptr(ctx, AC_UD_PUSH_CONSTANTS, user_sgpr_idx);
}
+ if (ctx->shader_info->info.num_inline_push_consts) {
+ set_loc_shader(ctx, AC_UD_INLINE_PUSH_CONSTANTS, user_sgpr_idx,
+ ctx->shader_info->info.num_inline_push_consts);
+ }
+
if (ctx->streamout_buffers) {
set_loc_shader_ptr(ctx, AC_UD_STREAMOUT_BUFFERS,
user_sgpr_idx);
switch (stage) {
case MESA_SHADER_COMPUTE:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
if (ctx->shader_info->info.cs.uses_grid_size) {
add_arg(&args, ARG_SGPR, ctx->ac.v3i32,
&ctx->abi.local_invocation_ids);
break;
case MESA_SHADER_VERTEX:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
+
declare_vs_specific_input_sgprs(ctx, stage, has_previous_stage,
previous_stage, &args);
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
+
declare_vs_specific_input_sgprs(ctx, stage,
has_previous_stage,
previous_stage, &args);
declare_vs_input_vgprs(ctx, &args);
} else {
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (needs_view_index)
}
break;
case MESA_SHADER_TESS_EVAL:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
if (needs_view_index)
add_arg(&args, ARG_SGPR, ctx->ac.i32,
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (previous_stage != MESA_SHADER_TESS_EVAL) {
declare_tes_input_vgprs(ctx, &args);
}
} else {
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (needs_view_index)
}
break;
case MESA_SHADER_FRAGMENT:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->abi.prim_mask);
add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->persp_sample);
if (has_previous_stage)
user_sgpr_idx = 0;
- set_global_input_locs(ctx, stage, has_previous_stage, previous_stage,
- &user_sgpr_info, desc_sets, &user_sgpr_idx);
+ set_global_input_locs(ctx, &user_sgpr_info, desc_sets, &user_sgpr_idx);
switch (stage) {
case MESA_SHADER_COMPUTE:
} else
stride = LLVMConstInt(ctx->ac.i32, layout->binding[binding].size, false);
- offset = ac_build_imad(&ctx->ac, index, stride,
- LLVMConstInt(ctx->ac.i32, base_offset, false));
+ offset = LLVMConstInt(ctx->ac.i32, base_offset, false);
- desc_ptr = ac_build_gep0(&ctx->ac, desc_ptr, offset);
+ if (layout->binding[binding].type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+ offset = ac_build_imad(&ctx->ac, index, stride, offset);
+ }
+
+ desc_ptr = LLVMBuildGEP(ctx->ac.builder, desc_ptr, &offset, 1, "");
desc_ptr = ac_cast_ptr(&ctx->ac, desc_ptr, ctx->ac.v4i32);
LLVMSetMetadata(desc_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
+ if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+ uint32_t desc_type = 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);
+
+ LLVMValueRef desc_components[4] = {
+ LLVMBuildPtrToInt(ctx->ac.builder, desc_ptr, ctx->ac.intptr, ""),
+ LLVMConstInt(ctx->ac.i32, S_008F04_BASE_ADDRESS_HI(ctx->options->address32_hi), false),
+ /* High limit to support variable sizes. */
+ LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
+ LLVMConstInt(ctx->ac.i32, desc_type, false),
+ };
+
+ return ac_build_gather_values(&ctx->ac, desc_components, 4);
+ }
+
return desc_ptr;
}
{
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
const unsigned location = var->data.location;
- const unsigned component = var->data.location_frac;
+ unsigned component = var->data.location_frac;
const bool is_patch = var->data.patch;
const bool is_compact = var->data.compact;
LLVMValueRef dw_addr;
}
param = shader_io_get_unique_index(location);
- if (location == VARYING_SLOT_CLIP_DIST0 &&
- is_compact && const_index > 3) {
- const_index -= 3;
- param++;
+ if ((location == VARYING_SLOT_CLIP_DIST0 || location == VARYING_SLOT_CLIP_DIST1) && is_compact) {
+ const_index += component;
+ component = 0;
+
+ if (const_index >= 4) {
+ const_index -= 4;
+ param++;
+ }
}
if (!is_patch) {
if (!is_tess_factor && writemask != 0xF)
ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, value, 1,
buf_addr, ctx->oc_lds,
- 4 * (base + chan), 1, 0, true, false);
+ 4 * (base + chan), 1, 0, false);
}
if (writemask == 0xF) {
ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, src, 4,
buf_addr, ctx->oc_lds,
- (base * 4), 1, 0, true, false);
+ (base * 4), 1, 0, false);
}
}
LLVMValueRef result;
unsigned param = shader_io_get_unique_index(location);
- if (location == VARYING_SLOT_CLIP_DIST0 && is_compact && const_index > 3) {
- const_index -= 3;
- param++;
+ if ((location == VARYING_SLOT_CLIP_DIST0 || location == VARYING_SLOT_CLIP_DIST1) && is_compact) {
+ const_index += component;
+ component = 0;
+ if (const_index >= 4) {
+ const_index -= 4;
+ param++;
+ }
}
buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index,
case 8:
sample_pos_offset = 7;
break;
- case 16:
- sample_pos_offset = 15;
- break;
default:
break;
}
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
LLVMValueRef result;
- LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false));
+ LLVMValueRef index = LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false);
+ LLVMValueRef ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ring_offsets, &index, 1, "");
ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
ac_array_in_const_addr_space(ctx->ac.v2f32), "");
ctx->gsvs_ring[stream],
out_val, 1,
voffset, ctx->gs2vs_offset, 0,
- 1, 1, true, true);
+ 1, 1, true);
}
}
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
LLVMValueRef result;
+ if (LLVMGetTypeKind(LLVMTypeOf(buffer_ptr)) != LLVMPointerTypeKind) {
+ /* Do not load the descriptor for inlined uniform blocks. */
+ return buffer_ptr;
+ }
+
LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
result = LLVMBuildLoad(ctx->ac.builder, buffer_ptr, "");
break;
case AC_DESC_SAMPLER:
type = ctx->ac.v4i32;
- if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
- offset += 64;
+ if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
+ offset += radv_combined_image_descriptor_sampler_offset(binding);
+ }
type_size = 16;
break;
type = ctx->ac.v4i32;
type_size = 16;
break;
+ case AC_DESC_PLANE_0:
+ case AC_DESC_PLANE_1:
+ case AC_DESC_PLANE_2:
+ type = ctx->ac.v8i32;
+ type_size = 32;
+ offset += 32 * (desc_type - AC_DESC_PLANE_0);
+ break;
default:
unreachable("invalid desc_type\n");
}
assert(stride % type_size == 0);
- if (!index)
- index = ctx->ac.i32_0;
+ LLVMValueRef adjusted_index = index;
+ if (!adjusted_index)
+ adjusted_index = ctx->ac.i32_0;
- index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
+ adjusted_index = LLVMBuildMul(builder, adjusted_index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
- list = ac_build_gep0(&ctx->ac, list, LLVMConstInt(ctx->ac.i32, offset, 0));
+ LLVMValueRef val_offset = LLVMConstInt(ctx->ac.i32, offset, 0);
+ list = LLVMBuildGEP(builder, list, &val_offset, 1, "");
list = LLVMBuildPointerCast(builder, list,
ac_array_in_const32_addr_space(type), "");
- return ac_build_load_to_sgpr(&ctx->ac, list, index);
+ LLVMValueRef descriptor = ac_build_load_to_sgpr(&ctx->ac, list, adjusted_index);
+
+ /* 3 plane formats always have same size and format for plane 1 & 2, so
+ * use the tail from plane 1 so that we can store only the first 16 bytes
+ * of the last plane. */
+ if (desc_type == AC_DESC_PLANE_2) {
+ LLVMValueRef descriptor2 = radv_get_sampler_desc(abi, descriptor_set, base_index, constant_index, index, AC_DESC_PLANE_1,image, write, bindless);
+
+ LLVMValueRef components[8];
+ for (unsigned i = 0; i < 4; ++i)
+ components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor, i);
+
+ for (unsigned i = 4; i < 8; ++i)
+ components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor2, i);
+ descriptor = ac_build_gather_values(&ctx->ac, components, 8);
+ }
+
+ return descriptor;
}
/* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
LLVMValueRef c30 = LLVMConstInt(ctx->ac.i32, 30, 0);
+ alpha = LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.f32, "");
+
if (adjustment == RADV_ALPHA_ADJUST_SSCALED)
alpha = LLVMBuildFPToUI(ctx->ac.builder, alpha, ctx->ac.i32, "");
else
alpha = LLVMBuildSIToFP(ctx->ac.builder, alpha, ctx->ac.f32, "");
}
- return alpha;
+ return LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.i32, "");
+}
+
+static unsigned
+get_num_channels_from_data_format(unsigned data_format)
+{
+ switch (data_format) {
+ case V_008F0C_BUF_DATA_FORMAT_8:
+ case V_008F0C_BUF_DATA_FORMAT_16:
+ case V_008F0C_BUF_DATA_FORMAT_32:
+ return 1;
+ case V_008F0C_BUF_DATA_FORMAT_8_8:
+ case V_008F0C_BUF_DATA_FORMAT_16_16:
+ case V_008F0C_BUF_DATA_FORMAT_32_32:
+ return 2;
+ case V_008F0C_BUF_DATA_FORMAT_10_11_11:
+ case V_008F0C_BUF_DATA_FORMAT_11_11_10:
+ case V_008F0C_BUF_DATA_FORMAT_32_32_32:
+ return 3;
+ case V_008F0C_BUF_DATA_FORMAT_8_8_8_8:
+ case V_008F0C_BUF_DATA_FORMAT_10_10_10_2:
+ case V_008F0C_BUF_DATA_FORMAT_2_10_10_10:
+ case V_008F0C_BUF_DATA_FORMAT_16_16_16_16:
+ case V_008F0C_BUF_DATA_FORMAT_32_32_32_32:
+ return 4;
+ default:
+ break;
+ }
+
+ return 4;
+}
+
+static LLVMValueRef
+radv_fixup_vertex_input_fetches(struct radv_shader_context *ctx,
+ LLVMValueRef value,
+ unsigned num_channels,
+ bool is_float)
+{
+ LLVMValueRef zero = is_float ? ctx->ac.f32_0 : ctx->ac.i32_0;
+ LLVMValueRef one = is_float ? ctx->ac.f32_1 : ctx->ac.i32_1;
+ LLVMValueRef chan[4];
+
+ if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
+ unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
+
+ if (num_channels == 4 && num_channels == vec_size)
+ return value;
+
+ num_channels = MIN2(num_channels, vec_size);
+
+ for (unsigned i = 0; i < num_channels; i++)
+ chan[i] = ac_llvm_extract_elem(&ctx->ac, value, i);
+ } else {
+ if (num_channels) {
+ assert(num_channels == 1);
+ chan[0] = value;
+ }
+ }
+
+ for (unsigned i = num_channels; i < 4; i++) {
+ chan[i] = i == 3 ? one : zero;
+ chan[i] = ac_to_integer(&ctx->ac, chan[i]);
+ }
+
+ return ac_build_gather_values(&ctx->ac, chan, 4);
}
static void
unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
uint8_t input_usage_mask =
ctx->shader_info->info.vs.input_usage_mask[variable->data.location];
- unsigned num_channels = util_last_bit(input_usage_mask);
+ unsigned num_input_channels = util_last_bit(input_usage_mask);
variable->data.driver_location = variable->data.location * 4;
for (unsigned i = 0; i < attrib_count; ++i) {
LLVMValueRef output[4];
unsigned attrib_index = variable->data.location + i - VERT_ATTRIB_GENERIC0;
+ unsigned attrib_format = ctx->options->key.vs.vertex_attribute_formats[attrib_index];
+ unsigned data_format = attrib_format & 0x0f;
+ unsigned num_format = (attrib_format >> 4) & 0x07;
+ bool is_float = num_format != V_008F0C_BUF_NUM_FORMAT_UINT &&
+ num_format != V_008F0C_BUF_NUM_FORMAT_SINT;
if (ctx->options->key.vs.instance_rate_inputs & (1u << attrib_index)) {
uint32_t divisor = ctx->options->key.vs.instance_rate_divisors[attrib_index];
} else
buffer_index = LLVMBuildAdd(ctx->ac.builder, ctx->abi.vertex_id,
ctx->abi.base_vertex, "");
- t_offset = LLVMConstInt(ctx->ac.i32, attrib_index, false);
- t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
+ /* Adjust the number of channels to load based on the vertex
+ * attribute format.
+ */
+ unsigned num_format_channels = get_num_channels_from_data_format(data_format);
+ unsigned num_channels = MIN2(num_input_channels, num_format_channels);
+ unsigned attrib_binding = ctx->options->key.vs.vertex_attribute_bindings[attrib_index];
+ unsigned attrib_offset = ctx->options->key.vs.vertex_attribute_offsets[attrib_index];
+ unsigned attrib_stride = ctx->options->key.vs.vertex_attribute_strides[attrib_index];
+
+ if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) {
+ /* Always load, at least, 3 channels for formats that
+ * need to be shuffled because X<->Z.
+ */
+ num_channels = MAX2(num_channels, 3);
+ }
+
+ if (attrib_stride != 0 && attrib_offset > attrib_stride) {
+ LLVMValueRef buffer_offset =
+ LLVMConstInt(ctx->ac.i32,
+ attrib_offset / attrib_stride, false);
- input = ac_build_buffer_load_format(&ctx->ac, t_list,
+ buffer_index = LLVMBuildAdd(ctx->ac.builder,
buffer_index,
- ctx->ac.i32_0,
- num_channels, false, true);
+ buffer_offset, "");
- input = ac_build_expand_to_vec4(&ctx->ac, input, num_channels);
+ attrib_offset = attrib_offset % attrib_stride;
+ }
+
+ t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false);
+ t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
+
+ input = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
+ buffer_index,
+ LLVMConstInt(ctx->ac.i32, attrib_offset, false),
+ ctx->ac.i32_0, ctx->ac.i32_0,
+ num_channels,
+ data_format, num_format,
+ false, false, true);
+
+ if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) {
+ LLVMValueRef c[4];
+ c[0] = ac_llvm_extract_elem(&ctx->ac, input, 2);
+ c[1] = ac_llvm_extract_elem(&ctx->ac, input, 1);
+ c[2] = ac_llvm_extract_elem(&ctx->ac, input, 0);
+ c[3] = ac_llvm_extract_elem(&ctx->ac, input, 3);
+
+ input = ac_build_gather_values(&ctx->ac, c, 4);
+ }
+
+ input = radv_fixup_vertex_input_fetches(ctx, input, num_channels,
+ is_float);
for (unsigned chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
unsigned attr,
LLVMValueRef interp_param,
LLVMValueRef prim_mask,
+ bool float16,
LLVMValueRef result[4])
{
LLVMValueRef attr_number;
for (chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
- if (interp) {
+ if (interp && float16) {
+ result[chan] = ac_build_fs_interp_f16(&ctx->ac,
+ llvm_chan,
+ attr_number,
+ prim_mask, i, j);
+ } else if (interp) {
result[chan] = ac_build_fs_interp(&ctx->ac,
llvm_chan,
attr_number,
attr_number,
prim_mask);
result[chan] = LLVMBuildBitCast(ctx->ac.builder, result[chan], ctx->ac.i32, "");
- result[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, result[chan], LLVMTypeOf(interp_param), "");
+ result[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, result[chan], float16 ? ctx->ac.i16 : ctx->ac.i32, "");
+ }
+ }
+}
+
+static void mark_16bit_fs_input(struct radv_shader_context *ctx,
+ const struct glsl_type *type,
+ int location)
+{
+ if (glsl_type_is_scalar(type) || glsl_type_is_vector(type) || glsl_type_is_matrix(type)) {
+ unsigned attrib_count = glsl_count_attribute_slots(type, false);
+ if (glsl_type_is_16bit(type)) {
+ ctx->float16_shaded_mask |= ((1ull << attrib_count) - 1) << location;
+ }
+ } else if (glsl_type_is_array(type)) {
+ unsigned stride = glsl_count_attribute_slots(glsl_get_array_element(type), false);
+ for (unsigned i = 0; i < glsl_get_length(type); ++i) {
+ mark_16bit_fs_input(ctx, glsl_get_array_element(type), location + i * stride);
+ }
+ } else {
+ assert(glsl_type_is_struct_or_ifc(type));
+ for (unsigned i = 0; i < glsl_get_length(type); i++) {
+ mark_16bit_fs_input(ctx, glsl_get_struct_field(type, i), location);
+ location += glsl_count_attribute_slots(glsl_get_struct_field(type, i), false);
}
}
}
uint64_t mask;
variable->data.driver_location = idx * 4;
+
+
+ if (variable->data.compact) {
+ unsigned component_count = variable->data.location_frac +
+ glsl_get_length(variable->type);
+ attrib_count = (component_count + 3) / 4;
+ } else
+ mark_16bit_fs_input(ctx, variable->type, idx);
+
mask = ((1ull << attrib_count) - 1) << variable->data.location;
- if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT) {
+ if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT ||
+ glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT16 ||
+ glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_STRUCT) {
unsigned interp_type;
if (variable->data.sample)
interp_type = INTERP_SAMPLE;
interp = lookup_interp_param(&ctx->abi, variable->data.interpolation, interp_type);
}
- bool is_16bit = glsl_type_is_16bit(variable->type);
- LLVMTypeRef type = is_16bit ? ctx->ac.i16 : ctx->ac.i32;
if (interp == NULL)
- interp = LLVMGetUndef(type);
+ interp = LLVMGetUndef(ctx->ac.i32);
for (unsigned i = 0; i < attrib_count; ++i)
ctx->inputs[ac_llvm_reg_index_soa(idx + i, 0)] = interp;
- if (idx == VARYING_SLOT_CLIP_DIST0) {
- /* Do not account for the number of components inside the array
- * of clip/cull distances because this might wrongly set other
- * bits like primitive ID or layer.
- */
- mask = 1ull << VARYING_SLOT_CLIP_DIST0;
- }
-
ctx->input_mask |= mask;
}
if (i >= VARYING_SLOT_VAR0 || i == VARYING_SLOT_PNTC ||
i == VARYING_SLOT_PRIMITIVE_ID || i == VARYING_SLOT_LAYER) {
interp_param = *inputs;
- interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask,
+ bool float16 = (ctx->float16_shaded_mask >> i) & 1;
+ interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask, float16,
inputs);
if (LLVMIsUndef(interp_param))
ctx->shader_info->fs.flat_shaded_mask |= 1u << index;
+ if (float16)
+ ctx->shader_info->fs.float16_shaded_mask |= 1u << index;
+ if (i >= VARYING_SLOT_VAR0)
+ ctx->abi.fs_input_attr_indices[i - VARYING_SLOT_VAR0] = index;
++index;
} else if (i == VARYING_SLOT_CLIP_DIST0) {
int length = ctx->shader_info->info.ps.num_input_clips_culls;
interp_param = *inputs;
interp_fs_input(ctx, index, interp_param,
- ctx->abi.prim_mask, inputs);
+ ctx->abi.prim_mask, false, inputs);
++index;
}
} else if (i == VARYING_SLOT_POS) {
if (stage == MESA_SHADER_TESS_CTRL)
return;
+ if (variable->data.compact) {
+ unsigned component_count = variable->data.location_frac +
+ glsl_get_length(variable->type);
+ attrib_count = (component_count + 3) / 4;
+ }
+
mask_attribs = ((1ull << attrib_count) - 1) << idx;
if (stage == MESA_SHADER_VERTEX ||
stage == MESA_SHADER_TESS_EVAL ||
ctx->shader_info->tes.outinfo.cull_dist_mask = (1 << shader->info.cull_distance_array_size) - 1;
ctx->shader_info->tes.outinfo.cull_dist_mask <<= shader->info.clip_distance_array_size;
}
-
- mask_attribs = 1ull << idx;
}
}
if (is_16bit) {
for (unsigned chan = 0; chan < 4; chan++)
values[chan] = LLVMBuildZExt(ctx->ac.builder,
- values[chan],
+ ac_to_integer(&ctx->ac, values[chan]),
ctx->ac.i32, "");
}
break;
if (is_16bit) {
for (unsigned chan = 0; chan < 4; chan++)
values[chan] = LLVMBuildSExt(ctx->ac.builder,
- values[chan],
+ ac_to_integer(&ctx->ac, values[chan]),
ctx->ac.i32, "");
}
break;
} else
memcpy(&args->out[0], values, sizeof(values[0]) * 4);
- for (unsigned i = 0; i < 4; ++i) {
- if (!(args->enabled_channels & (1 << i)))
- continue;
-
+ for (unsigned i = 0; i < 4; ++i)
args->out[i] = ac_to_float(&ctx->ac, args->out[i]);
- }
}
static void
/* Get the first component. */
start = ffs(output->component_mask) - 1;
- /* Adjust the destination offset. */
- offset += start * 4;
-
/* Load the output as int. */
for (int i = 0; i < num_comps; i++) {
out[i] = ac_to_integer(&ctx->ac,
/* fall through */
case 4: /* as v4i32 */
vdata = ac_build_gather_values(&ctx->ac, out,
- util_next_power_of_two(num_comps));
+ !ac_has_vec3_support(ctx->ac.chip_class, false) ?
+ util_next_power_of_two(num_comps) :
+ num_comps);
break;
}
ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf],
vdata, num_comps, so_write_offsets[buf],
ctx->ac.i32_0, offset,
- 1, 1, true, false);
+ 1, 1, false);
}
static void
static void
handle_vs_outputs_post(struct radv_shader_context *ctx,
bool export_prim_id, bool export_layer_id,
+ bool export_clip_dists,
struct radv_vs_output_info *outinfo)
{
uint32_t param_count = 0;
memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
sizeof(outinfo->vs_output_param_offset));
- if (ctx->output_mask & (1ull << VARYING_SLOT_CLIP_DIST0)) {
- unsigned output_usage_mask, length;
- LLVMValueRef slots[8];
- unsigned j;
-
- if (ctx->stage == MESA_SHADER_VERTEX &&
- !ctx->is_gs_copy_shader) {
- output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
- output_usage_mask =
- ctx->shader_info->info.tes.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- } else {
- assert(ctx->is_gs_copy_shader);
- output_usage_mask =
- ctx->shader_info->info.gs.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- }
+ for(unsigned location = VARYING_SLOT_CLIP_DIST0; location <= VARYING_SLOT_CLIP_DIST1; ++location) {
+ if (ctx->output_mask & (1ull << location)) {
+ unsigned output_usage_mask, length;
+ LLVMValueRef slots[4];
+ unsigned j;
+
+ if (ctx->stage == MESA_SHADER_VERTEX &&
+ !ctx->is_gs_copy_shader) {
+ output_usage_mask =
+ ctx->shader_info->info.vs.output_usage_mask[location];
+ } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ output_usage_mask =
+ ctx->shader_info->info.tes.output_usage_mask[location];
+ } else {
+ assert(ctx->is_gs_copy_shader);
+ output_usage_mask =
+ ctx->shader_info->info.gs.output_usage_mask[location];
+ }
- length = util_last_bit(output_usage_mask);
+ length = util_last_bit(output_usage_mask);
- i = VARYING_SLOT_CLIP_DIST0;
- for (j = 0; j < length; j++)
- slots[j] = ac_to_float(&ctx->ac, radv_load_output(ctx, i, j));
+ for (j = 0; j < length; j++)
+ slots[j] = ac_to_float(&ctx->ac, radv_load_output(ctx, location, j));
- for (i = length; i < 8; i++)
- slots[i] = LLVMGetUndef(ctx->ac.f32);
+ for (i = length; i < 4; i++)
+ slots[i] = LLVMGetUndef(ctx->ac.f32);
- if (length > 4) {
- target = V_008DFC_SQ_EXP_POS + 3;
- si_llvm_init_export_args(ctx, &slots[4], 0xf, target, &args);
+ target = V_008DFC_SQ_EXP_POS + 2 + (location - VARYING_SLOT_CLIP_DIST0);
+ si_llvm_init_export_args(ctx, &slots[0], 0xf, target, &args);
memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
- &args, sizeof(args));
- }
-
- target = V_008DFC_SQ_EXP_POS + 2;
- si_llvm_init_export_args(ctx, &slots[0], 0xf, target, &args);
- memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
- &args, sizeof(args));
+ &args, sizeof(args));
- /* Export the clip/cull distances values to the next stage. */
- radv_export_param(ctx, param_count, &slots[0], 0xf);
- outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0] = param_count++;
- if (length > 4) {
- radv_export_param(ctx, param_count, &slots[4], 0xf);
- outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1] = param_count++;
+ if (export_clip_dists) {
+ /* Export the clip/cull distances values to the next stage. */
+ radv_export_param(ctx, param_count, &slots[0], 0xf);
+ outinfo->vs_output_param_offset[location] = param_count++;
+ }
}
}
LLVMValueRef lds_base = NULL;
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
- unsigned output_usage_mask;
int param_index;
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
- if (ctx->stage == MESA_SHADER_VERTEX) {
- output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[i];
- } else {
- assert(ctx->stage == MESA_SHADER_TESS_EVAL);
- output_usage_mask =
- ctx->shader_info->info.tes.output_usage_mask[i];
- }
-
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
param_index = shader_io_get_unique_index(i);
- max_output_written = MAX2(param_index + (length > 4), max_output_written);
+ max_output_written = MAX2(param_index, max_output_written);
}
outinfo->esgs_itemsize = (max_output_written + 1) * 16;
LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
unsigned output_usage_mask;
int param_index;
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
ctx->shader_info->info.tes.output_usage_mask[i];
}
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
param_index = shader_io_get_unique_index(i);
if (lds_base) {
"");
}
- for (j = 0; j < length; j++) {
+ for (j = 0; j < 4; j++) {
if (!(output_usage_mask & (1 << j)))
continue;
out_val, 1,
NULL, ctx->es2gs_offset,
(4 * param_index + j) * 4,
- 1, 1, true, true);
+ 1, 1, true);
}
}
}
vertex_dw_stride, "");
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
- unsigned output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[i];
LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
int param = shader_io_get_unique_index(i);
LLVMValueRef dw_addr = LLVMBuildAdd(ctx->ac.builder, base_dw_addr,
LLVMConstInt(ctx->ac.i32, param * 4, false),
"");
- for (unsigned j = 0; j < length; j++) {
+ for (unsigned j = 0; j < 4; j++) {
LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, out_ptr[j], "");
value = ac_to_integer(&ctx->ac, value);
value = LLVMBuildZExtOrBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
LLVMConstInt(ctx->ac.i32, 4 * stride, false), "");
unsigned tf_offset = 0;
- if (ctx->options->chip_class <= VI) {
+ if (ctx->options->chip_class <= GFX8) {
ac_nir_build_if(&inner_if_ctx, ctx,
LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
rel_patch_id, ctx->ac.i32_0, ""));
ac_build_buffer_store_dword(&ctx->ac, buffer,
LLVMConstInt(ctx->ac.i32, 0x80000000, false),
1, ctx->ac.i32_0, tf_base,
- 0, 1, 0, true, false);
+ 0, 1, 0, false);
tf_offset += 4;
ac_nir_build_endif(&inner_if_ctx);
/* Store the tessellation factors. */
ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
MIN2(stride, 4), byteoffset, tf_base,
- tf_offset, 1, 0, true, false);
+ tf_offset, 1, 0, false);
if (vec1)
ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
stride - 4, byteoffset, tf_base,
- 16 + tf_offset, 1, 0, true, false);
+ 16 + tf_offset, 1, 0, false);
//store to offchip for TES to read - only if TES reads them
if (ctx->options->key.tcs.tes_reads_tess_factors) {
ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, outer_vec,
outer_comps, tf_outer_offset,
- ctx->oc_lds, 0, 1, 0, true, false);
+ ctx->oc_lds, 0, 1, 0, false);
if (inner_comps) {
param_inner = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER);
tf_inner_offset = get_tcs_tes_buffer_address(ctx, NULL,
ac_build_gather_values(&ctx->ac, inner, inner_comps);
ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, inner_vec,
inner_comps, tf_inner_offset,
- ctx->oc_lds, 0, 1, 0, true, false);
+ ctx->oc_lds, 0, 1, 0, false);
}
}
ac_nir_build_endif(&if_ctx);
else
handle_vs_outputs_post(ctx, ctx->options->key.vs.export_prim_id,
ctx->options->key.vs.export_layer_id,
+ ctx->options->key.vs.export_clip_dists,
&ctx->shader_info->vs.outinfo);
break;
case MESA_SHADER_FRAGMENT:
else
handle_vs_outputs_post(ctx, ctx->options->key.tes.export_prim_id,
ctx->options->key.tes.export_layer_id,
+ ctx->options->key.tes.export_clip_dists,
&ctx->shader_info->tes.outinfo);
break;
default:
static void
ac_setup_rings(struct radv_shader_context *ctx)
{
- if (ctx->options->chip_class <= VI &&
+ if (ctx->options->chip_class <= GFX8 &&
(ctx->stage == MESA_SHADER_GEOMETRY ||
ctx->options->key.vs.as_es || ctx->options->key.tes.as_es)) {
unsigned ring = ctx->stage == MESA_SHADER_GEOMETRY ? RING_ESGS_GS
stride = 4 * num_components * ctx->gs_max_out_vertices;
- /* Limit on the stride field for <= CIK. */
+ /* Limit on the stride field for <= GFX7. */
assert(stride < (1 << 14));
ring = LLVMBuildBitCast(ctx->ac.builder,
}
}
-static unsigned
-ac_nir_get_max_workgroup_size(enum chip_class chip_class,
- const struct nir_shader *nir)
+unsigned
+radv_nir_get_max_workgroup_size(enum chip_class chip_class,
+ const struct nir_shader *nir)
{
switch (nir->info.stage) {
case MESA_SHADER_TESS_CTRL:
- return chip_class >= CIK ? 128 : 64;
+ return chip_class >= GFX7 ? 128 : 64;
case MESA_SHADER_GEOMETRY:
return chip_class >= GFX9 ? 128 : 64;
case MESA_SHADER_COMPUTE:
memset(shader_info, 0, sizeof(*shader_info));
+ radv_nir_shader_info_init(&shader_info->info);
+
for(int i = 0; i < shader_count; ++i)
radv_nir_shader_info_pass(shaders[i], options, &shader_info->info);
ctx.max_workgroup_size = 0;
for (int i = 0; i < shader_count; ++i) {
ctx.max_workgroup_size = MAX2(ctx.max_workgroup_size,
- ac_nir_get_max_workgroup_size(ctx.options->chip_class,
+ radv_nir_get_max_workgroup_size(ctx.options->chip_class,
shaders[i]));
}
ctx.abi.load_sampler_desc = radv_get_sampler_desc;
ctx.abi.load_resource = radv_load_resource;
ctx.abi.clamp_shadow_reference = false;
- ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9;
+ ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9 && HAVE_LLVM < 0x800;
+
+ /* Because the new raw/struct atomic intrinsics are buggy with LLVM 8,
+ * we fallback to the old intrinsics for atomic buffer image operations
+ * and thus we need to apply the indexing workaround...
+ */
+ ctx.abi.gfx9_stride_size_workaround_for_atomic = ctx.ac.chip_class == GFX9 && HAVE_LLVM < 0x900;
if (shader_count >= 2)
ac_init_exec_full_mask(&ctx.ac);
- if (ctx.ac.chip_class == GFX9 &&
+ if ((ctx.ac.family == CHIP_VEGA10 ||
+ ctx.ac.family == CHIP_RAVEN) &&
shaders[shader_count - 1]->info.stage == MESA_SHADER_TESS_CTRL)
ac_nir_fixup_ls_hs_input_vgprs(&ctx);
* - Floating-point output modifiers would be ignored by the hw.
* - Some opcodes don't support denormals, such as v_mad_f32. We would
* have to stop using those.
- * - SI & CI would be very slow.
+ * - GFX6 & GFX7 would be very slow.
*/
config->float_mode |= V_00B028_FP_64_DENORMS;
}
radv_emit_streamout(ctx, stream);
if (stream == 0) {
- handle_vs_outputs_post(ctx, false, false,
+ handle_vs_outputs_post(ctx, false, false, true,
&ctx->shader_info->vs.outinfo);
}
projects / mesa.git / blobdiff