#include "sid.h"
#define MAX_GLOBAL_BUFFERS 20
-#if HAVE_LLVM < 0x0305
-#define NUM_USER_SGPRS 2
-#else
-/* XXX: Even though we don't pass the scratch buffer via user sgprs any more
- * LLVM still expects that we specify 4 USER_SGPRS so it can remain compatible
- * with older mesa. */
-#define NUM_USER_SGPRS 4
-#endif
struct si_compute {
struct si_context *ctx;
program->shader.binary.global_symbol_offsets[i];
unsigned scratch_bytes_needed;
- si_shader_binary_read_config(sctx->screen,
- &program->shader, offset);
- scratch_bytes_needed = program->shader.scratch_bytes_per_wave;
+ si_shader_binary_read_config(&program->shader.binary,
+ &program->shader.config, offset);
+ scratch_bytes_needed = program->shader.config.scratch_bytes_per_wave;
scratch_bytes = MAX2(scratch_bytes, scratch_bytes_needed);
}
if (scratch_bytes == 0)
return;
- program->shader.scratch_bo = (struct r600_resource*)
+ program->shader.scratch_bo =
si_resource_create_custom(sctx->b.b.screen,
PIPE_USAGE_DEFAULT,
scratch_bytes * scratch_waves);
* to the maximum bytes needed, so it can compute the stride
* correctly.
*/
- program->shader.scratch_bytes_per_wave = scratch_bytes;
+ program->shader.config.scratch_bytes_per_wave = scratch_bytes;
/* Patch the shader with the scratch buffer address. */
si_shader_apply_scratch_relocs(sctx,
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);
+ si_compile_llvm(sctx->screen, &program->kernels[i].binary,
+ &program->kernels[i].config, sctx->tm,
+ mod, &sctx->b.debug, TGSI_PROCESSOR_COMPUTE);
+ si_shader_binary_upload(sctx->screen, &program->kernels[i]);
LLVMDisposeModule(mod);
}
}
#else
- radeon_elf_read(code, header->num_bytes, &program->shader.binary, true);
+ radeon_elf_read(code, header->num_bytes, &program->shader.binary);
/* init_scratch_buffer patches the shader code with the scratch address,
* so we need to call it before si_shader_binary_read() which uploads
* the shader code to the GPU.
*/
init_scratch_buffer(sctx, program);
- si_shader_binary_read(sctx->screen, &program->shader, &program->shader.binary);
+ si_shader_binary_read(sctx->screen, &program->shader.binary,
+ &program->shader.config, &sctx->b.debug,
+ TGSI_PROCESSOR_COMPUTE);
+ si_shader_binary_upload(sctx->screen, &program->shader);
#endif
program->input_buffer = si_resource_create_custom(sctx->b.b.screen,
uint32_t pc, const void *input)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
+ struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
struct si_compute *program = sctx->cs_shader_state.program;
struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
struct r600_resource *input_buffer = program->input_buffer;
uint64_t kernel_args_va;
uint64_t scratch_buffer_va = 0;
uint64_t shader_va;
- unsigned arg_user_sgpr_count = NUM_USER_SGPRS;
unsigned i;
struct si_shader *shader = &program->shader;
unsigned lds_blocks;
radeon_emit(cs, 0x80000000);
radeon_emit(cs, 0x80000000);
- sctx->b.flags |= SI_CONTEXT_INV_TC_L1 |
- SI_CONTEXT_INV_TC_L2 |
+ sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
+ SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_ICACHE |
- SI_CONTEXT_INV_KCACHE |
+ SI_CONTEXT_INV_SMEM_L1 |
SI_CONTEXT_FLUSH_WITH_INV_L2 |
SI_CONTEXT_FLAG_COMPUTE;
- si_emit_cache_flush(&sctx->b, NULL);
+ si_emit_cache_flush(sctx, NULL);
pm4->compute_pkt = true;
#if HAVE_LLVM >= 0x0306
/* Read the config information */
- si_shader_binary_read_config(sctx->screen, shader, pc);
+ si_shader_binary_read_config(&shader->binary, &shader->config, 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 + 8 /* For scratch va */;
- kernel_args = sctx->b.ws->buffer_map(input_buffer->cs_buf,
- sctx->b.rings.gfx.cs, PIPE_TRANSFER_WRITE);
+ kernel_args = sctx->b.ws->buffer_map(input_buffer->buf,
+ sctx->b.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];
memcpy(kernel_args + (num_work_size_bytes / 4), input, program->input_size);
- if (shader->scratch_bytes_per_wave > 0) {
+ if (shader->config.scratch_bytes_per_wave > 0) {
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,
- shader->scratch_bytes_per_wave *
+ shader->config.scratch_bytes_per_wave,
+ shader->config.scratch_bytes_per_wave *
num_waves_for_scratch);
- si_pm4_add_bo(pm4, shader->scratch_bo,
- RADEON_USAGE_READWRITE,
- RADEON_PRIO_SHADER_RESOURCE_RW);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
+ shader->scratch_bo,
+ RADEON_USAGE_READWRITE,
+ RADEON_PRIO_SCRATCH_BUFFER);
scratch_buffer_va = shader->scratch_bo->gpu_address;
}
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, input_buffer, RADEON_USAGE_READ,
- RADEON_PRIO_SHADER_DATA);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, input_buffer,
+ RADEON_USAGE_READ, RADEON_PRIO_CONST_BUFFER);
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));
+ | S_008F04_STRIDE(shader->config.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, RADEON_PRIO_SHADER_RESOURCE_RW);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, buffer,
+ RADEON_USAGE_READWRITE,
+ RADEON_PRIO_COMPUTE_GLOBAL);
}
/* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
#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);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, shader->bo,
+ RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
+ si_pm4_set_reg(pm4, R_00B830_COMPUTE_PGM_LO, shader_va >> 8);
si_pm4_set_reg(pm4, R_00B834_COMPUTE_PGM_HI, shader_va >> 40);
- si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1,
- /* We always use at least 3 VGPRS, these come from
- * TIDIG_COMP_CNT.
- * XXX: The compiler should account for this.
- */
- S_00B848_VGPRS((MAX2(3, shader->num_vgprs) - 1) / 4)
- /* We always use at least 4 + arg_user_sgpr_count. The 4 extra
- * sgprs are from TGID_X_EN, TGID_Y_EN, TGID_Z_EN, TG_SIZE_EN
- * XXX: The compiler should account for this.
- */
- | S_00B848_SGPRS(((MAX2(4 + arg_user_sgpr_count,
- shader->num_sgprs)) - 1) / 8))
- ;
+ si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1, shader->config.rsrc1);
- lds_blocks = shader->lds_size;
+ lds_blocks = shader->config.lds_size;
/* XXX: We are over allocating LDS. For SI, the shader reports LDS in
* blocks of 256 bytes, so if there are 4 bytes lds allocated in
* the shader and 4 bytes allocated by the state tracker, then
assert(lds_blocks <= 0xFF);
- si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2,
- 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_00B84C_TGID_Z_EN(1)
- | S_00B84C_TG_SIZE_EN(1)
- | S_00B84C_TIDIG_COMP_CNT(2)
- | S_00B84C_LDS_SIZE(lds_blocks)
- | S_00B84C_EXCP_EN(0))
- ;
+ shader->config.rsrc2 &= C_00B84C_LDS_SIZE;
+ shader->config.rsrc2 |= S_00B84C_LDS_SIZE(lds_blocks);
+
+ si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2, shader->config.rsrc2);
si_pm4_set_reg(pm4, R_00B854_COMPUTE_RESOURCE_LIMITS, 0);
si_pm4_set_reg(pm4, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0,
* COMPUTE_PGM_RSRC2.SCRATCH_EN is enabled.
*/
S_00B860_WAVES(num_waves_for_scratch)
- | S_00B860_WAVESIZE(shader->scratch_bytes_per_wave >> 10))
+ | S_00B860_WAVESIZE(shader->config.scratch_bytes_per_wave >> 10))
;
si_pm4_cmd_begin(pm4, PKT3_DISPATCH_DIRECT);
si_pm4_free_state(sctx, pm4, ~0);
sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
- SI_CONTEXT_INV_TC_L1 |
- SI_CONTEXT_INV_TC_L2 |
+ SI_CONTEXT_INV_VMEM_L1 |
+ SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_ICACHE |
- SI_CONTEXT_INV_KCACHE |
+ SI_CONTEXT_INV_SMEM_L1 |
SI_CONTEXT_FLAG_COMPUTE;
- si_emit_cache_flush(&sctx->b, NULL);
+ si_emit_cache_flush(sctx, NULL);
}
if (program->kernels) {
for (int i = 0; i < program->num_kernels; i++){
if (program->kernels[i].bo){
- si_shader_destroy(ctx, &program->kernels[i]);
+ si_shader_destroy(&program->kernels[i]);
}
}
FREE(program->kernels);
FREE(program->shader.binary.config);
FREE(program->shader.binary.rodata);
FREE(program->shader.binary.global_symbol_offsets);
- si_shader_destroy(ctx, &program->shader);
+ si_shader_destroy(&program->shader);
#endif
pipe_resource_reference(