*/
#include "util/u_memory.h"
+#include "radeon/r600_pipe_common.h"
+#include "radeon/radeon_elf_util.h"
+#include "radeon/radeon_llvm_util.h"
-#include "../radeon/r600_cs.h"
+#include "radeon/r600_cs.h"
#include "si_pipe.h"
#include "si_shader.h"
#include "sid.h"
-#include "radeon_llvm_util.h"
-
#define MAX_GLOBAL_BUFFERS 20
+#if HAVE_LLVM < 0x0305
+#define NUM_USER_SGPRS 2
+#else
+#define NUM_USER_SGPRS 4
+#endif
-struct si_pipe_compute {
+struct si_compute {
struct si_context *ctx;
unsigned local_size;
unsigned private_size;
unsigned input_size;
- unsigned num_kernels;
- struct si_pipe_shader *kernels;
+ struct radeon_shader_binary binary;
+ struct si_shader program;
unsigned num_user_sgprs;
- struct pipe_resource *global_buffers[MAX_GLOBAL_BUFFERS];
+ struct r600_resource *input_buffer;
+ struct pipe_resource *global_buffers[MAX_GLOBAL_BUFFERS];
+#if HAVE_LLVM < 0x0306
+ unsigned num_kernels;
+ struct si_shader *kernels;
LLVMContextRef llvm_ctx;
+#endif
};
static void *si_create_compute_state(
const struct pipe_compute_state *cso)
{
struct si_context *sctx = (struct si_context *)ctx;
- struct si_pipe_compute *program =
- CALLOC_STRUCT(si_pipe_compute);
+ struct si_compute *program = CALLOC_STRUCT(si_compute);
const struct pipe_llvm_program_header *header;
- const unsigned char *code;
- unsigned i;
-
- program->llvm_ctx = LLVMContextCreate();
+ const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
- program->num_kernels = radeon_llvm_get_num_kernels(program->llvm_ctx, code,
- header->num_bytes);
- program->kernels = CALLOC(sizeof(struct si_pipe_shader),
- program->num_kernels);
- for (i = 0; i < program->num_kernels; i++) {
- LLVMModuleRef mod = radeon_llvm_get_kernel_module(program->llvm_ctx, i,
- code, header->num_bytes);
- si_compile_llvm(sctx, &program->kernels[i], mod);
- LLVMDisposeModule(mod);
+#if HAVE_LLVM < 0x0306
+ {
+ unsigned i;
+ program->llvm_ctx = LLVMContextCreate();
+ program->num_kernels = radeon_llvm_get_num_kernels(program->llvm_ctx,
+ code, header->num_bytes);
+ program->kernels = CALLOC(sizeof(struct si_shader),
+ program->num_kernels);
+ for (i = 0; i < program->num_kernels; i++) {
+ LLVMModuleRef mod = radeon_llvm_get_kernel_module(program->llvm_ctx, i,
+ code, header->num_bytes);
+ si_compile_llvm(sctx->screen, &program->kernels[i], mod);
+ LLVMDisposeModule(mod);
+ }
}
+#else
+
+ radeon_elf_read(code, header->num_bytes, &program->binary, true);
+ si_shader_binary_read(sctx->screen, &program->program, &program->binary);
+
+#endif
+ program->input_buffer = si_resource_create_custom(sctx->b.b.screen,
+ PIPE_USAGE_IMMUTABLE, program->input_size);
return program;
}
static void si_bind_compute_state(struct pipe_context *ctx, void *state)
{
struct si_context *sctx = (struct si_context*)ctx;
- sctx->cs_shader_state.program = (struct si_pipe_compute*)state;
+ sctx->cs_shader_state.program = (struct si_compute*)state;
}
static void si_set_global_binding(
{
unsigned i;
struct si_context *sctx = (struct si_context*)ctx;
- struct si_pipe_compute *program = sctx->cs_shader_state.program;
+ struct si_compute *program = sctx->cs_shader_state.program;
if (!resources) {
for (i = first; i < first + n; i++) {
- program->global_buffers[i] = NULL;
+ pipe_resource_reference(&program->global_buffers[i], NULL);
}
return;
}
for (i = first; i < first + n; i++) {
uint64_t va;
uint32_t offset;
- program->global_buffers[i] = resources[i];
- va = r600_resource_va(ctx->screen, resources[i]);
+ pipe_resource_reference(&program->global_buffers[i], resources[i]);
+ va = r600_resource(resources[i])->gpu_address;
offset = util_le32_to_cpu(*handles[i]);
va += offset;
va = util_cpu_to_le64(va);
}
}
+/**
+ * This function computes the value for R_00B860_COMPUTE_TMPRING_SIZE.WAVES
+ * /p block_layout is the number of threads in each work group.
+ * /p grid layout is the number of work groups.
+ */
+static unsigned compute_num_waves_for_scratch(
+ const struct radeon_info *info,
+ const uint *block_layout,
+ const uint *grid_layout)
+{
+ unsigned num_sh = MAX2(info->max_sh_per_se, 1);
+ unsigned num_se = MAX2(info->max_se, 1);
+ unsigned num_blocks = 1;
+ unsigned threads_per_block = 1;
+ unsigned waves_per_block;
+ unsigned waves_per_sh;
+ unsigned waves;
+ unsigned scratch_waves;
+ unsigned i;
+
+ for (i = 0; i < 3; i++) {
+ threads_per_block *= block_layout[i];
+ num_blocks *= grid_layout[i];
+ }
+
+ waves_per_block = align(threads_per_block, 64) / 64;
+ waves = waves_per_block * num_blocks;
+ waves_per_sh = align(waves, num_sh * num_se) / (num_sh * num_se);
+ scratch_waves = waves_per_sh * num_sh * num_se;
+
+ if (waves_per_block > waves_per_sh) {
+ scratch_waves = waves_per_block * num_sh * num_se;
+ }
+
+ return scratch_waves;
+}
+
static void si_launch_grid(
struct pipe_context *ctx,
const uint *block_layout, const uint *grid_layout,
uint32_t pc, const void *input)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct si_pipe_compute *program = sctx->cs_shader_state.program;
+ struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
+ struct si_compute *program = sctx->cs_shader_state.program;
struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
- struct r600_resource *kernel_args_buffer = NULL;
+ struct r600_resource *input_buffer = program->input_buffer;
unsigned kernel_args_size;
unsigned num_work_size_bytes = 36;
uint32_t kernel_args_offset = 0;
uint32_t *kernel_args;
uint64_t kernel_args_va;
+ uint64_t scratch_buffer_va = 0;
uint64_t shader_va;
- unsigned arg_user_sgpr_count = 2;
+ unsigned arg_user_sgpr_count = NUM_USER_SGPRS;
unsigned i;
- struct si_pipe_shader *shader = &program->kernels[pc];
+ struct si_shader *shader = &program->program;
unsigned lds_blocks;
+ unsigned num_waves_for_scratch;
- pm4->compute_pkt = true;
- si_cmd_context_control(pm4);
+#if HAVE_LLVM < 0x0306
+ shader = &program->kernels[pc];
+#endif
- si_pm4_cmd_begin(pm4, PKT3_EVENT_WRITE);
- si_pm4_cmd_add(pm4, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH) |
- EVENT_INDEX(0x7) |
- EVENT_WRITE_INV_L2);
- si_pm4_cmd_end(pm4, false);
- si_pm4_inval_texture_cache(pm4);
- si_pm4_inval_shader_cache(pm4);
- si_cmd_surface_sync(pm4, pm4->cp_coher_cntl);
+ radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0) | PKT3_SHADER_TYPE_S(1));
+ radeon_emit(cs, 0x80000000);
+ radeon_emit(cs, 0x80000000);
+
+ sctx->b.flags |= R600_CONTEXT_INV_TEX_CACHE |
+ R600_CONTEXT_INV_SHADER_CACHE |
+ R600_CONTEXT_INV_CONST_CACHE |
+ R600_CONTEXT_FLUSH_WITH_INV_L2 |
+ R600_CONTEXT_FLAG_COMPUTE;
+ si_emit_cache_flush(&sctx->b, NULL);
+
+ pm4->compute_pkt = true;
+
+#if HAVE_LLVM >= 0x0306
+ /* Read the config information */
+ si_shader_binary_read_config(&program->binary, &program->program, pc);
+#endif
/* Upload the kernel arguments */
/* The extra num_work_size_bytes are for work group / work item size information */
- kernel_args_size = program->input_size + num_work_size_bytes;
- kernel_args = MALLOC(kernel_args_size);
+ kernel_args_size = program->input_size + num_work_size_bytes + 8 /* For scratch va */;
+
+ kernel_args = sctx->b.ws->buffer_map(input_buffer->cs_buf,
+ sctx->b.rings.gfx.cs, PIPE_TRANSFER_WRITE);
for (i = 0; i < 3; i++) {
kernel_args[i] = grid_layout[i];
kernel_args[i + 3] = grid_layout[i] * block_layout[i];
kernel_args[i + 6] = block_layout[i];
}
+ num_waves_for_scratch = compute_num_waves_for_scratch(
+ &sctx->screen->b.info, block_layout, grid_layout);
+
memcpy(kernel_args + (num_work_size_bytes / 4), input, program->input_size);
- si_upload_const_buffer(sctx, &kernel_args_buffer, (uint8_t*)kernel_args,
- kernel_args_size, &kernel_args_offset);
- kernel_args_va = r600_resource_va(ctx->screen,
- (struct pipe_resource*)kernel_args_buffer);
+ if (shader->scratch_bytes_per_wave > 0) {
+ unsigned scratch_bytes = shader->scratch_bytes_per_wave *
+ num_waves_for_scratch;
+
+ COMPUTE_DBG(sctx->screen, "Waves: %u; Scratch per wave: %u bytes; "
+ "Total Scratch: %u bytes\n", num_waves_for_scratch,
+ shader->scratch_bytes_per_wave, scratch_bytes);
+ if (!shader->scratch_bo) {
+ shader->scratch_bo = (struct r600_resource*)
+ si_resource_create_custom(sctx->b.b.screen,
+ PIPE_USAGE_DEFAULT, scratch_bytes);
+ }
+ scratch_buffer_va = shader->scratch_bo->gpu_address;
+ si_pm4_add_bo(pm4, shader->scratch_bo,
+ RADEON_USAGE_READWRITE,
+ RADEON_PRIO_SHADER_RESOURCE_RW);
+
+ }
+
+ for (i = 0; i < (kernel_args_size / 4); i++) {
+ COMPUTE_DBG(sctx->screen, "input %u : %u\n", i,
+ kernel_args[i]);
+ }
+
+ sctx->b.ws->buffer_unmap(input_buffer->cs_buf);
+
+ kernel_args_va = input_buffer->gpu_address;
kernel_args_va += kernel_args_offset;
- si_pm4_add_bo(pm4, kernel_args_buffer, RADEON_USAGE_READ);
+ si_pm4_add_bo(pm4, input_buffer, RADEON_USAGE_READ,
+ RADEON_PRIO_SHADER_DATA);
si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0, kernel_args_va);
si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 4, S_008F04_BASE_ADDRESS_HI (kernel_args_va >> 32) | S_008F04_STRIDE(0));
+ si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 8, scratch_buffer_va);
+ si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 12,
+ S_008F04_BASE_ADDRESS_HI(scratch_buffer_va >> 32)
+ | S_008F04_STRIDE(shader->scratch_bytes_per_wave / 64));
si_pm4_set_reg(pm4, R_00B810_COMPUTE_START_X, 0);
si_pm4_set_reg(pm4, R_00B814_COMPUTE_START_Y, 0);
if (!buffer) {
continue;
}
- si_pm4_add_bo(pm4, buffer, RADEON_USAGE_READWRITE);
+ si_pm4_add_bo(pm4, buffer, RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RESOURCE_RW);
}
/* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
0x190 /* Default value */);
}
- shader_va = r600_resource_va(ctx->screen, (void *)shader->bo);
- si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ);
+ shader_va = shader->bo->gpu_address;
+
+#if HAVE_LLVM >= 0x0306
+ shader_va += pc;
+#endif
+ si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_DATA);
si_pm4_set_reg(pm4, R_00B830_COMPUTE_PGM_LO, (shader_va >> 8) & 0xffffffff);
si_pm4_set_reg(pm4, R_00B834_COMPUTE_PGM_HI, shader_va >> 40);
assert(lds_blocks <= 0xFF);
si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2,
- S_00B84C_SCRATCH_EN(0)
+ S_00B84C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0)
| S_00B84C_USER_SGPR(arg_user_sgpr_count)
| S_00B84C_TGID_X_EN(1)
| S_00B84C_TGID_Y_EN(1)
| S_00B85C_SH1_CU_EN(0xffff /* Default value */))
;
+ num_waves_for_scratch =
+ MIN2(num_waves_for_scratch,
+ 32 * sctx->screen->b.info.max_compute_units);
+ si_pm4_set_reg(pm4, R_00B860_COMPUTE_TMPRING_SIZE,
+ /* The maximum value for WAVES is 32 * num CU.
+ * If you program this value incorrectly, the GPU will hang if
+ * COMPUTE_PGM_RSRC2.SCRATCH_EN is enabled.
+ */
+ S_00B860_WAVES(num_waves_for_scratch)
+ | S_00B860_WAVESIZE(shader->scratch_bytes_per_wave >> 10))
+ ;
+
si_pm4_cmd_begin(pm4, PKT3_DISPATCH_DIRECT);
si_pm4_cmd_add(pm4, grid_layout[0]); /* Thread groups DIM_X */
si_pm4_cmd_add(pm4, grid_layout[1]); /* Thread groups DIM_Y */
si_pm4_cmd_add(pm4, 1); /* DISPATCH_INITIATOR */
si_pm4_cmd_end(pm4, false);
- si_pm4_cmd_begin(pm4, PKT3_EVENT_WRITE);
- si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(0x4)));
- si_pm4_cmd_end(pm4, false);
-
- si_pm4_inval_texture_cache(pm4);
- si_pm4_inval_shader_cache(pm4);
- si_cmd_surface_sync(pm4, pm4->cp_coher_cntl);
-
si_pm4_emit(sctx, pm4);
#if 0
}
#endif
- FREE(pm4);
- FREE(kernel_args);
+ si_pm4_free_state(sctx, pm4, ~0);
+
+ sctx->b.flags |= R600_CONTEXT_CS_PARTIAL_FLUSH |
+ R600_CONTEXT_INV_TEX_CACHE |
+ R600_CONTEXT_INV_SHADER_CACHE |
+ R600_CONTEXT_INV_CONST_CACHE |
+ R600_CONTEXT_FLAG_COMPUTE;
+ si_emit_cache_flush(&sctx->b, NULL);
}
static void si_delete_compute_state(struct pipe_context *ctx, void* state){
- struct si_pipe_compute *program = (struct si_pipe_compute *)state;
+ struct si_compute *program = (struct si_compute *)state;
if (!state) {
return;
}
+#if HAVE_LLVM < 0x0306
if (program->kernels) {
+ for (int i = 0; i < program->num_kernels; i++){
+ if (program->kernels[i].bo){
+ si_shader_destroy(ctx, &program->kernels[i]);
+ }
+ }
FREE(program->kernels);
}
if (program->llvm_ctx){
LLVMContextDispose(program->llvm_ctx);
}
+#else
+ si_shader_destroy(ctx, &program->program);
+#endif
+
+ pipe_resource_reference(
+ (struct pipe_resource **)&program->input_buffer, NULL);
- //And then free the program itself.
+ FREE(program->binary.code);
+ FREE(program->binary.config);
+ FREE(program->binary.rodata);
FREE(program);
}