GL_ARB_bindless_texture DONE (nvc0, radeonsi)
GL_ARB_cl_event not started
- GL_ARB_compute_variable_group_size DONE (nvc0, radeonsi)
+ GL_ARB_compute_variable_group_size DONE (i965/gen7+, nvc0, radeonsi)
GL_ARB_ES3_2_compatibility DONE (i965/gen8+, radeonsi, virgl)
GL_ARB_fragment_shader_interlock DONE (i965)
GL_ARB_gpu_shader_int64 DONE (i965/gen8+, nvc0, radeonsi, softpipe, llvmpipe)
+GL_ARB_compute_variable_group_size on i965.
GL_EXT_texture_shadow_lod on radeonsi.
GL_NV_copy_image on all gallium drivers.
VK_KHR_shader_non_semantic_info on Intel, RADV.
brw->compute.num_work_groups_bo = NULL;
brw->compute.num_work_groups = num_groups;
+ brw->compute.group_size = NULL;
ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
brw_dispatch_compute_common(ctx);
brw->compute.num_work_groups_bo = bo;
brw->compute.num_work_groups_offset = indirect;
brw->compute.num_work_groups = indirect_group_counts;
+ brw->compute.group_size = NULL;
+ ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
+
+ brw_dispatch_compute_common(ctx);
+}
+
+static void
+brw_dispatch_compute_group_size(struct gl_context *ctx,
+ const GLuint *num_groups,
+ const GLuint *group_size)
+{
+ struct brw_context *brw = brw_context(ctx);
+
+ brw->compute.num_work_groups_bo = NULL;
+ brw->compute.num_work_groups = num_groups;
+ brw->compute.group_size = group_size;
ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
brw_dispatch_compute_common(ctx);
{
functions->DispatchCompute = brw_dispatch_compute;
functions->DispatchComputeIndirect = brw_dispatch_compute_indirect;
+ functions->DispatchComputeGroupSize = brw_dispatch_compute_group_size;
}
ctx->Const.MaxComputeWorkGroupSize[2] = max_invocations;
ctx->Const.MaxComputeWorkGroupInvocations = max_invocations;
ctx->Const.MaxComputeSharedMemorySize = 64 * 1024;
+
+ /* Constants used for ARB_compute_variable_group_size. The compiler will
+ * use the maximum to decide which SIMDs can be used. If we top this like
+ * max_invocations, that would prevent SIMD8 / SIMD16 to be considered.
+ *
+ * TODO: To avoid the trade off above between having the lower maximum
+ * vs. always using SIMD32, keep all three shader variants (for each SIMD)
+ * and select a suitable one at dispatch time.
+ */
+ if (devinfo->gen >= 7) {
+ const uint32_t max_var_invocations =
+ (max_threads >= 64 ? 8 : (max_threads >= 32 ? 16 : 32)) * max_threads;
+ assert(max_var_invocations >= 512);
+ ctx->Const.MaxComputeVariableGroupSize[0] = max_var_invocations;
+ ctx->Const.MaxComputeVariableGroupSize[1] = max_var_invocations;
+ ctx->Const.MaxComputeVariableGroupSize[2] = max_var_invocations;
+ ctx->Const.MaxComputeVariableGroupInvocations = max_var_invocations;
+ }
}
/**
struct brw_bo *num_work_groups_bo;
GLintptr num_work_groups_offset;
const GLuint *num_work_groups;
+ /**
+ * This is only used alongside ARB_compute_variable_group_size when the
+ * local work group size is variable, otherwise it's NULL.
+ */
+ const GLuint *group_size;
} compute;
struct {
#include "brw_program.h"
#include "compiler/glsl/ir_uniform.h"
+uint32_t
+brw_cs_group_size(const struct brw_context *brw)
+{
+ assert(brw->cs.base.prog_data);
+ struct brw_cs_prog_data *cs_prog_data =
+ brw_cs_prog_data(brw->cs.base.prog_data);
+
+ if (brw->compute.group_size) {
+ /* With ARB_compute_variable_group_size the group size is set at
+ * dispatch time, so we can't use the one provided by the compiler.
+ */
+ return brw->compute.group_size[0] *
+ brw->compute.group_size[1] *
+ brw->compute.group_size[2];
+ } else {
+ return cs_prog_data->local_size[0] *
+ cs_prog_data->local_size[1] *
+ cs_prog_data->local_size[2];
+ }
+}
+
static void
assign_cs_binding_table_offsets(const struct gen_device_info *devinfo,
const struct gl_program *prog,
struct brw_cs_prog_data prog_data;
bool start_busy = false;
double start_time = 0;
+ struct gl_context *gl_ctx = &brw->ctx;
nir_shader *nir = nir_shader_clone(mem_ctx, cp->program.nir);
memset(&prog_data, 0, sizeof(prog_data));
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
st_index = brw_get_shader_time_index(brw, &cp->program, ST_CS, true);
+ /* If the work group size is variable we set it to the maximum here since
+ * the actual size is not known until the dispatch command is issued.
+ */
+ if (nir->info.cs.local_size_variable) {
+ prog_data.uses_variable_group_size = true;
+ nir->info.cs.max_variable_local_size =
+ gl_ctx->Const.MaxComputeWorkGroupSize[2];
+ } else {
+ prog_data.uses_variable_group_size = false;
+ }
+
char *error_str;
program = brw_compile_cs(brw->screen->compiler, brw, mem_ctx, key,
&prog_data, nir, st_index, NULL, &error_str);
extern "C" {
#endif
+uint32_t
+brw_cs_group_size(const struct brw_context *brw);
+
void
brw_upload_cs_prog(struct brw_context *brw);
*/
#include "brw_context.h"
+#include "brw_cs.h"
#include "brw_state.h"
#include "brw_defines.h"
#include "brw_program.h"
return f_as_u32(ctx->TessCtrlProgram.patch_default_inner_level[0]);
} else if (param == BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y) {
return f_as_u32(ctx->TessCtrlProgram.patch_default_inner_level[1]);
+ } else if (param >= BRW_PARAM_BUILTIN_WORK_GROUP_SIZE_X &&
+ param <= BRW_PARAM_BUILTIN_WORK_GROUP_SIZE_Z) {
+ unsigned i = param - BRW_PARAM_BUILTIN_WORK_GROUP_SIZE_X;
+ return brw->compute.group_size[i];
} else {
unreachable("Invalid param builtin");
}
/* XXX: Should this happen somewhere before to get our state flag set? */
_mesa_load_state_parameters(ctx, prog->Parameters);
+ const unsigned threads =
+ DIV_ROUND_UP(brw_cs_group_size(brw), cs_prog_data->simd_size);
const unsigned push_const_size =
- brw_cs_push_const_total_size(cs_prog_data, cs_prog_data->threads);
+ brw_cs_push_const_total_size(cs_prog_data, threads);
+
if (push_const_size == 0) {
stage_state->push_const_size = 0;
return;
}
if (cs_prog_data->push.per_thread.size > 0) {
- for (unsigned t = 0; t < cs_prog_data->threads; t++) {
+ for (unsigned t = 0; t < threads; t++) {
unsigned dst =
8 * (cs_prog_data->push.per_thread.regs * t +
cs_prog_data->push.cross_thread.regs);
#include "genX_boilerplate.h"
#include "brw_context.h"
+#include "brw_cs.h"
#include "brw_draw.h"
#include "brw_multisample_state.h"
#include "brw_state.h"
struct brw_cs_prog_data *cs_prog_data = brw_cs_prog_data(prog_data);
const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ const unsigned threads =
+ DIV_ROUND_UP(brw_cs_group_size(brw), cs_prog_data->simd_size);
+
+ if (!cs_prog_data->uses_variable_group_size)
+ assert(cs_prog_data->threads == threads);
+
if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
brw_emit_buffer_surface_state(
brw, &stage_state->surf_offset[
vfe.URBEntryAllocationSize = GEN_GEN >= 8 ? 2 : 0;
const uint32_t vfe_curbe_allocation =
- ALIGN(cs_prog_data->push.per_thread.regs * cs_prog_data->threads +
+ ALIGN(cs_prog_data->push.per_thread.regs * threads +
cs_prog_data->push.cross_thread.regs, 2);
vfe.CURBEAllocationSize = vfe_curbe_allocation;
}
const unsigned push_const_size =
- brw_cs_push_const_total_size(cs_prog_data, cs_prog_data->threads);
+ brw_cs_push_const_total_size(cs_prog_data, threads);
if (push_const_size > 0) {
brw_batch_emit(brw, GENX(MEDIA_CURBE_LOAD), curbe) {
curbe.CURBETotalDataLength = ALIGN(push_const_size, 64);
DIV_ROUND_UP(CLAMP(stage_state->sampler_count, 0, 16), 4),
.BindingTablePointer = stage_state->bind_bo_offset,
.ConstantURBEntryReadLength = cs_prog_data->push.per_thread.regs,
- .NumberofThreadsinGPGPUThreadGroup = cs_prog_data->threads,
+ .NumberofThreadsinGPGPUThreadGroup = threads,
.SharedLocalMemorySize = encode_slm_size(GEN_GEN,
prog_data->total_shared),
.BarrierEnable = cs_prog_data->uses_barrier,
if (indirect)
prepare_indirect_gpgpu_walker(brw);
+ const unsigned group_size = brw_cs_group_size(brw);
const unsigned simd_size = prog_data->simd_size;
- unsigned group_size = prog_data->local_size[0] *
- prog_data->local_size[1] * prog_data->local_size[2];
+ unsigned thread_width_max = DIV_ROUND_UP(group_size, simd_size);
uint32_t right_mask = 0xffffffffu >> (32 - simd_size);
const unsigned right_non_aligned = group_size & (simd_size - 1);
ggw.SIMDSize = prog_data->simd_size / 16;
ggw.ThreadDepthCounterMaximum = 0;
ggw.ThreadHeightCounterMaximum = 0;
- ggw.ThreadWidthCounterMaximum = prog_data->threads - 1;
+ ggw.ThreadWidthCounterMaximum = thread_width_max - 1;
ggw.ThreadGroupIDXDimension = num_groups[0];
ggw.ThreadGroupIDYDimension = num_groups[1];
ggw.ThreadGroupIDZDimension = num_groups[2];
ctx->Extensions.ARB_ES3_1_compatibility =
devinfo->gen >= 8 || devinfo->is_haswell;
ctx->Extensions.NV_compute_shader_derivatives = true;
+ ctx->Extensions.ARB_compute_variable_group_size = true;
}
if (can_do_predicate_writes(brw->screen)) {
projects / mesa.git / commitdiff