2
* Copyright © Microsoft Corporation
4
* Permission is hereby granted, free of charge, to any person obtaining a
5
* copy of this software and associated documentation files (the "Software"),
6
* to deal in the Software without restriction, including without limitation
7
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
8
* and/or sell copies of the Software, and to permit persons to whom the
9
* Software is furnished to do so, subject to the following conditions:
11
* The above copyright notice and this permission notice (including the next
12
* paragraph) shall be included in all copies or substantial portions of the
15
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24
#include "d3d12_compute_transforms.h"
25
#include "d3d12_nir_passes.h"
26
#include "d3d12_query.h"
27
#include "d3d12_screen.h"
30
#include "nir_builder.h"
32
#include "util/u_memory.h"
35
get_indirect_draw_base_vertex_transform(const nir_shader_compiler_options *options, const d3d12_compute_transform_key *args)
37
nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_COMPUTE, options, "TransformIndirectDrawBaseVertex");
39
if (args->base_vertex.dynamic_count) {
40
nir_variable *count_ubo = nir_variable_create(b.shader, nir_var_mem_ubo,
41
glsl_uint_type(), "in_count");
42
count_ubo->data.driver_location = 0;
45
nir_variable *input_ssbo = nir_variable_create(b.shader, nir_var_mem_ssbo,
46
glsl_array_type(glsl_uint_type(), 0, 0), "input");
47
nir_variable *output_ssbo = nir_variable_create(b.shader, nir_var_mem_ssbo,
48
input_ssbo->type, "output");
49
input_ssbo->data.driver_location = 0;
50
output_ssbo->data.driver_location = 1;
52
nir_ssa_def *draw_id = nir_channel(&b, nir_load_global_invocation_id(&b, 32), 0);
53
if (args->base_vertex.dynamic_count) {
54
nir_ssa_def *count = nir_load_ubo(&b, 1, 32, nir_imm_int(&b, 1), nir_imm_int(&b, 0),
55
(gl_access_qualifier)0, 4, 0, 0, 4);
56
nir_push_if(&b, nir_ilt(&b, draw_id, count));
59
nir_variable *stride_ubo = NULL;
60
nir_ssa_def *in_stride_offset_and_base_drawid = d3d12_get_state_var(&b, D3D12_STATE_VAR_TRANSFORM_GENERIC0, "d3d12_Stride",
61
glsl_uvec4_type(), &stride_ubo);
62
nir_ssa_def *in_offset = nir_iadd(&b, nir_channel(&b, in_stride_offset_and_base_drawid, 1),
63
nir_imul(&b, nir_channel(&b, in_stride_offset_and_base_drawid, 0), draw_id));
64
nir_ssa_def *in_data0 = nir_load_ssbo(&b, 4, 32, nir_imm_int(&b, 0), in_offset, (gl_access_qualifier)0, 4, 0);
66
nir_ssa_def *in_data1 = NULL;
67
nir_ssa_def *base_vertex = NULL, *base_instance = NULL;
68
if (args->base_vertex.indexed) {
69
nir_ssa_def *in_offset1 = nir_iadd(&b, in_offset, nir_imm_int(&b, 16));
70
in_data1 = nir_load_ssbo(&b, 1, 32, nir_imm_int(&b, 0), in_offset1, (gl_access_qualifier)0, 4, 0);
71
base_vertex = nir_channel(&b, in_data0, 3);
72
base_instance = in_data1;
74
base_vertex = nir_channel(&b, in_data0, 2);
75
base_instance = nir_channel(&b, in_data0, 3);
78
/* 4 additional uints for base vertex, base instance, draw ID, and a bool for indexed draw */
79
unsigned out_stride = sizeof(uint32_t) * ((args->base_vertex.indexed ? 5 : 4) + 4);
81
nir_ssa_def *out_offset = nir_imul(&b, draw_id, nir_imm_int(&b, out_stride));
82
nir_ssa_def *out_data0 = nir_vec4(&b, base_vertex, base_instance,
83
nir_iadd(&b, draw_id, nir_channel(&b, in_stride_offset_and_base_drawid, 2)),
84
nir_imm_int(&b, args->base_vertex.indexed ? -1 : 0));
85
nir_ssa_def *out_data1 = in_data0;
87
nir_store_ssbo(&b, out_data0, nir_imm_int(&b, 1), out_offset, 0xf, (gl_access_qualifier)0, 4, 0);
88
nir_store_ssbo(&b, out_data1, nir_imm_int(&b, 1), nir_iadd(&b, out_offset, nir_imm_int(&b, 16)),
89
(1u << out_data1->num_components) - 1, (gl_access_qualifier)0, 4, 0);
90
if (args->base_vertex.indexed)
91
nir_store_ssbo(&b, in_data1, nir_imm_int(&b, 1), nir_iadd(&b, out_offset, nir_imm_int(&b, 32)), 1, (gl_access_qualifier)0, 4, 0);
93
if (args->base_vertex.dynamic_count)
96
nir_validate_shader(b.shader, "creation");
97
b.shader->info.num_ssbos = 2;
98
b.shader->info.num_ubos = (args->base_vertex.dynamic_count ? 1 : 0);
103
static struct nir_shader *
104
get_fake_so_buffer_copy_back(const nir_shader_compiler_options *options, const d3d12_compute_transform_key *key)
106
nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_COMPUTE, options, "FakeSOBufferCopyBack");
108
nir_variable *output_so_data_var = nir_variable_create(b.shader, nir_var_mem_ssbo,
109
glsl_array_type(glsl_uint_type(), 0, 0), "output_data");
110
nir_variable *input_so_data_var = nir_variable_create(b.shader, nir_var_mem_ssbo, output_so_data_var->type, "input_data");
111
output_so_data_var->data.driver_location = 0;
112
input_so_data_var->data.driver_location = 1;
114
/* UBO is [fake SO filled size, fake SO vertex count, 1, 1, original SO filled size] */
115
nir_variable *input_ubo = nir_variable_create(b.shader, nir_var_mem_ubo,
116
glsl_array_type(glsl_uint_type(), 5, 0), "input_ubo");
117
input_ubo->data.driver_location = 0;
119
nir_ssa_def *original_so_filled_size = nir_load_ubo(&b, 1, 32, nir_imm_int(&b, 0), nir_imm_int(&b, 4 * sizeof(uint32_t)),
120
(gl_access_qualifier)0, 4, 0, 4 * sizeof(uint32_t), 4);
122
nir_variable *state_var = nullptr;
123
nir_ssa_def *fake_so_multiplier = d3d12_get_state_var(&b, D3D12_STATE_VAR_TRANSFORM_GENERIC0, "fake_so_multiplier", glsl_uint_type(), &state_var);
125
nir_ssa_def *vertex_offset = nir_imul(&b, nir_imm_int(&b, key->fake_so_buffer_copy_back.stride),
126
nir_channel(&b, nir_load_global_invocation_id(&b, 32), 0));
128
nir_ssa_def *output_offset_base = nir_iadd(&b, original_so_filled_size, vertex_offset);
129
nir_ssa_def *input_offset_base = nir_imul(&b, vertex_offset, fake_so_multiplier);
131
for (unsigned i = 0; i < key->fake_so_buffer_copy_back.num_ranges; ++i) {
132
auto& output = key->fake_so_buffer_copy_back.ranges[i];
133
assert(output.size % 4 == 0 && output.offset % 4 == 0);
134
nir_ssa_def *field_offset = nir_imm_int(&b, output.offset);
135
nir_ssa_def *output_offset = nir_iadd(&b, output_offset_base, field_offset);
136
nir_ssa_def *input_offset = nir_iadd(&b, input_offset_base, field_offset);
138
for (unsigned loaded = 0; loaded < output.size; loaded += 16) {
139
unsigned to_load = MIN2(output.size, 16);
140
unsigned components = to_load / 4;
141
nir_ssa_def *loaded_data = nir_load_ssbo(&b, components, 32, nir_imm_int(&b, 1),
142
nir_iadd(&b, input_offset, nir_imm_int(&b, loaded)), (gl_access_qualifier)0, 4, 0);
143
nir_store_ssbo(&b, loaded_data, nir_imm_int(&b, 0),
144
nir_iadd(&b, output_offset, nir_imm_int(&b, loaded)), (1u << components) - 1, (gl_access_qualifier)0, 4, 0);
148
nir_validate_shader(b.shader, "creation");
149
b.shader->info.num_ssbos = 2;
150
b.shader->info.num_ubos = 1;
155
static struct nir_shader *
156
get_fake_so_buffer_vertex_count(const nir_shader_compiler_options *options)
158
nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_COMPUTE, options, "FakeSOBufferVertexCount");
160
nir_variable_create(b.shader, nir_var_mem_ssbo, glsl_array_type(glsl_uint_type(), 0, 0), "fake_so");
161
nir_ssa_def *fake_buffer_filled_size = nir_load_ssbo(&b, 1, 32, nir_imm_int(&b, 0), nir_imm_int(&b, 0), (gl_access_qualifier)0, 4, 0);
163
nir_variable *real_so_var = nir_variable_create(b.shader, nir_var_mem_ssbo,
164
glsl_array_type(glsl_uint_type(), 0, 0), "real_so");
165
real_so_var->data.driver_location = 1;
166
nir_ssa_def *real_buffer_filled_size = nir_load_ssbo(&b, 1, 32, nir_imm_int(&b, 1), nir_imm_int(&b, 0), (gl_access_qualifier)0, 4, 0);
168
nir_variable *state_var = nullptr;
169
nir_ssa_def *state_var_data = d3d12_get_state_var(&b, D3D12_STATE_VAR_TRANSFORM_GENERIC0, "state_var", glsl_uvec4_type(), &state_var);
170
nir_ssa_def *stride = nir_channel(&b, state_var_data, 0);
171
nir_ssa_def *fake_so_multiplier = nir_channel(&b, state_var_data, 1);
173
nir_ssa_def *real_so_bytes_added = nir_idiv(&b, fake_buffer_filled_size, fake_so_multiplier);
174
nir_ssa_def *vertex_count = nir_idiv(&b, real_so_bytes_added, stride);
175
nir_ssa_def *to_write_to_fake_buffer = nir_vec4(&b, vertex_count, nir_imm_int(&b, 1), nir_imm_int(&b, 1), real_buffer_filled_size);
176
nir_store_ssbo(&b, to_write_to_fake_buffer, nir_imm_int(&b, 0), nir_imm_int(&b, 4), 0xf, (gl_access_qualifier)0, 4, 0);
178
nir_ssa_def *updated_filled_size = nir_iadd(&b, real_buffer_filled_size, real_so_bytes_added);
179
nir_store_ssbo(&b, updated_filled_size, nir_imm_int(&b, 1), nir_imm_int(&b, 0), 1, (gl_access_qualifier)0, 4, 0);
181
nir_validate_shader(b.shader, "creation");
182
b.shader->info.num_ssbos = 2;
183
b.shader->info.num_ubos = 0;
188
static struct nir_shader *
189
get_draw_auto(const nir_shader_compiler_options *options)
191
nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_COMPUTE, options, "DrawAuto");
193
nir_variable_create(b.shader, nir_var_mem_ssbo, glsl_array_type(glsl_uint_type(), 0, 0), "ssbo");
194
nir_ssa_def *buffer_filled_size = nir_load_ssbo(&b, 1, 32, nir_imm_int(&b, 0), nir_imm_int(&b, 0), (gl_access_qualifier)0, 4, 0);
196
nir_variable *state_var = nullptr;
197
nir_ssa_def *state_var_data = d3d12_get_state_var(&b, D3D12_STATE_VAR_TRANSFORM_GENERIC0, "state_var", glsl_uvec4_type(), &state_var);
198
nir_ssa_def *stride = nir_channel(&b, state_var_data, 0);
199
nir_ssa_def *vb_offset = nir_channel(&b, state_var_data, 1);
201
nir_ssa_def *vb_bytes = nir_bcsel(&b, nir_ilt(&b, vb_offset, buffer_filled_size),
202
nir_isub(&b, buffer_filled_size, vb_offset), nir_imm_int(&b, 0));
204
nir_ssa_def *vertex_count = nir_idiv(&b, vb_bytes, stride);
205
nir_ssa_def *to_write = nir_vec4(&b, vertex_count, nir_imm_int(&b, 1), nir_imm_int(&b, 0), nir_imm_int(&b, 0));
206
nir_store_ssbo(&b, to_write, nir_imm_int(&b, 0), nir_imm_int(&b, 4), 0xf, (gl_access_qualifier)0, 4, 0);
208
nir_validate_shader(b.shader, "creation");
209
b.shader->info.num_ssbos = 1;
210
b.shader->info.num_ubos = 0;
215
static struct nir_shader *
216
create_compute_transform(const nir_shader_compiler_options *options, const d3d12_compute_transform_key *key)
219
case d3d12_compute_transform_type::base_vertex:
220
return get_indirect_draw_base_vertex_transform(options, key);
221
case d3d12_compute_transform_type::fake_so_buffer_copy_back:
222
return get_fake_so_buffer_copy_back(options, key);
223
case d3d12_compute_transform_type::fake_so_buffer_vertex_count:
224
return get_fake_so_buffer_vertex_count(options);
225
case d3d12_compute_transform_type::draw_auto:
226
return get_draw_auto(options);
228
unreachable("Invalid transform");
232
struct compute_transform
234
d3d12_compute_transform_key key;
235
d3d12_shader_selector *shader;
238
d3d12_shader_selector *
239
d3d12_get_compute_transform(struct d3d12_context *ctx, const d3d12_compute_transform_key *key)
241
struct hash_entry *entry = _mesa_hash_table_search(ctx->compute_transform_cache, key);
243
compute_transform *data = (compute_transform *)MALLOC(sizeof(compute_transform));
247
const nir_shader_compiler_options *options = &d3d12_screen(ctx->base.screen)->nir_options;
249
memcpy(&data->key, key, sizeof(*key));
250
nir_shader *s = create_compute_transform(options, key);
255
struct pipe_compute_state shader_args = { PIPE_SHADER_IR_NIR, s };
256
data->shader = d3d12_create_compute_shader(ctx, &shader_args);
263
data->shader->is_variant = true;
264
entry = _mesa_hash_table_insert(ctx->compute_transform_cache, &data->key, data);
268
return ((struct compute_transform *)entry->data)->shader;
272
hash_compute_transform_key(const void *key)
274
return _mesa_hash_data(key, sizeof(struct d3d12_compute_transform_key));
278
equals_compute_transform_key(const void *a, const void *b)
280
return memcmp(a, b, sizeof(struct d3d12_compute_transform_key)) == 0;
284
d3d12_compute_transform_cache_init(struct d3d12_context *ctx)
286
ctx->compute_transform_cache = _mesa_hash_table_create(NULL,
287
hash_compute_transform_key,
288
equals_compute_transform_key);
292
delete_entry(struct hash_entry *entry)
294
struct compute_transform *data = (struct compute_transform *)entry->data;
295
d3d12_shader_free(data->shader);
300
d3d12_compute_transform_cache_destroy(struct d3d12_context *ctx)
302
_mesa_hash_table_destroy(ctx->compute_transform_cache, delete_entry);
306
d3d12_save_compute_transform_state(struct d3d12_context *ctx, d3d12_compute_transform_save_restore *save)
308
if (ctx->current_predication)
309
ctx->cmdlist->SetPredication(nullptr, 0, D3D12_PREDICATION_OP_EQUAL_ZERO);
311
memset(save, 0, sizeof(*save));
312
save->cs = ctx->compute_state;
314
pipe_resource_reference(&save->cbuf0.buffer, ctx->cbufs[PIPE_SHADER_COMPUTE][1].buffer);
315
save->cbuf0 = ctx->cbufs[PIPE_SHADER_COMPUTE][1];
317
for (unsigned i = 0; i < ARRAY_SIZE(save->ssbos); ++i) {
318
pipe_resource_reference(&save->ssbos[i].buffer, ctx->ssbo_views[PIPE_SHADER_COMPUTE][i].buffer);
319
save->ssbos[i] = ctx->ssbo_views[PIPE_SHADER_COMPUTE][i];
324
d3d12_restore_compute_transform_state(struct d3d12_context *ctx, d3d12_compute_transform_save_restore *save)
326
ctx->base.bind_compute_state(&ctx->base, save->cs);
328
ctx->base.set_constant_buffer(&ctx->base, PIPE_SHADER_COMPUTE, 1, true, &save->cbuf0);
329
ctx->base.set_shader_buffers(&ctx->base, PIPE_SHADER_COMPUTE, 0, ARRAY_SIZE(save->ssbos), save->ssbos, (1u << ARRAY_SIZE(save->ssbos)) - 1);
331
if (ctx->current_predication)
332
d3d12_enable_predication(ctx);