#include <sys/errno.h>
#include "main/condrender.h"
-#include "main/glheader.h"
#include "main/mtypes.h"
#include "main/state.h"
#include "brw_context.h"
#include "brw_draw.h"
#include "brw_state.h"
#include "intel_batchbuffer.h"
+#include "intel_buffer_objects.h"
#include "brw_defines.h"
static void
-brw_emit_gpgpu_walker(struct brw_context *brw, const GLuint *num_groups)
+prepare_indirect_gpgpu_walker(struct brw_context *brw)
+{
+ GLintptr indirect_offset = brw->compute.num_work_groups_offset;
+ drm_intel_bo *bo = brw->compute.num_work_groups_bo;
+
+ brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMX, bo,
+ I915_GEM_DOMAIN_VERTEX, 0,
+ indirect_offset + 0);
+ brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMY, bo,
+ I915_GEM_DOMAIN_VERTEX, 0,
+ indirect_offset + 4);
+ brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMZ, bo,
+ I915_GEM_DOMAIN_VERTEX, 0,
+ indirect_offset + 8);
+
+ if (brw->gen > 7)
+ return;
+
+ /* Clear upper 32-bits of SRC0 and all 64-bits of SRC1 */
+ BEGIN_BATCH(7);
+ OUT_BATCH(MI_LOAD_REGISTER_IMM | (7 - 2));
+ OUT_BATCH(MI_PREDICATE_SRC0 + 4);
+ OUT_BATCH(0u);
+ OUT_BATCH(MI_PREDICATE_SRC1 + 0);
+ OUT_BATCH(0u);
+ OUT_BATCH(MI_PREDICATE_SRC1 + 4);
+ OUT_BATCH(0u);
+ ADVANCE_BATCH();
+
+ /* Load compute_dispatch_indirect_x_size into SRC0 */
+ brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ indirect_offset + 0);
+
+ /* predicate = (compute_dispatch_indirect_x_size == 0); */
+ BEGIN_BATCH(1);
+ OUT_BATCH(GEN7_MI_PREDICATE |
+ MI_PREDICATE_LOADOP_LOAD |
+ MI_PREDICATE_COMBINEOP_SET |
+ MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
+ ADVANCE_BATCH();
+
+ /* Load compute_dispatch_indirect_y_size into SRC0 */
+ brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ indirect_offset + 4);
+
+ /* predicate |= (compute_dispatch_indirect_y_size == 0); */
+ BEGIN_BATCH(1);
+ OUT_BATCH(GEN7_MI_PREDICATE |
+ MI_PREDICATE_LOADOP_LOAD |
+ MI_PREDICATE_COMBINEOP_OR |
+ MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
+ ADVANCE_BATCH();
+
+ /* Load compute_dispatch_indirect_z_size into SRC0 */
+ brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ indirect_offset + 8);
+
+ /* predicate |= (compute_dispatch_indirect_z_size == 0); */
+ BEGIN_BATCH(1);
+ OUT_BATCH(GEN7_MI_PREDICATE |
+ MI_PREDICATE_LOADOP_LOAD |
+ MI_PREDICATE_COMBINEOP_OR |
+ MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
+ ADVANCE_BATCH();
+
+ /* predicate = !predicate; */
+ BEGIN_BATCH(1);
+ OUT_BATCH(GEN7_MI_PREDICATE |
+ MI_PREDICATE_LOADOP_LOADINV |
+ MI_PREDICATE_COMBINEOP_OR |
+ MI_PREDICATE_COMPAREOP_FALSE);
+ ADVANCE_BATCH();
+}
+
+static void
+brw_emit_gpgpu_walker(struct brw_context *brw)
{
const struct brw_cs_prog_data *prog_data = brw->cs.prog_data;
+ const GLuint *num_groups = brw->compute.num_work_groups;
+ uint32_t indirect_flag;
+
+ if (brw->compute.num_work_groups_bo == NULL) {
+ indirect_flag = 0;
+ } else {
+ indirect_flag =
+ GEN7_GPGPU_INDIRECT_PARAMETER_ENABLE |
+ (brw->gen == 7 ? GEN7_GPGPU_PREDICATE_ENABLE : 0);
+ prepare_indirect_gpgpu_walker(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 =
(group_size + simd_size - 1) / simd_size;
- uint32_t right_mask = (1u << simd_size) - 1;
+ uint32_t right_mask = 0xffffffffu >> (32 - simd_size);
const unsigned right_non_aligned = group_size & (simd_size - 1);
if (right_non_aligned != 0)
right_mask >>= (simd_size - right_non_aligned);
uint32_t dwords = brw->gen < 8 ? 11 : 15;
BEGIN_BATCH(dwords);
- OUT_BATCH(GPGPU_WALKER << 16 | (dwords - 2));
+ OUT_BATCH(GPGPU_WALKER << 16 | (dwords - 2) | indirect_flag);
OUT_BATCH(0);
if (brw->gen >= 8) {
OUT_BATCH(0); /* Indirect Data Length */
static void
-brw_dispatch_compute(struct gl_context *ctx, const GLuint *num_groups)
+brw_dispatch_compute_common(struct gl_context *ctx)
{
struct brw_context *brw = brw_context(ctx);
int estimated_buffer_space_needed;
brw->no_batch_wrap = true;
brw_upload_compute_state(brw);
- brw_emit_gpgpu_walker(brw, num_groups);
+ brw_emit_gpgpu_walker(brw);
brw->no_batch_wrap = false;
*/
}
+static void
+brw_dispatch_compute(struct gl_context *ctx, const GLuint *num_groups) {
+ struct brw_context *brw = brw_context(ctx);
+
+ brw->compute.num_work_groups_bo = NULL;
+ brw->compute.num_work_groups = num_groups;
+ ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
+
+ brw_dispatch_compute_common(ctx);
+}
+
+static void
+brw_dispatch_compute_indirect(struct gl_context *ctx, GLintptr indirect)
+{
+ struct brw_context *brw = brw_context(ctx);
+ static const GLuint indirect_group_counts[3] = { 0, 0, 0 };
+ struct gl_buffer_object *indirect_buffer = ctx->DispatchIndirectBuffer;
+ drm_intel_bo *bo =
+ intel_bufferobj_buffer(brw,
+ intel_buffer_object(indirect_buffer),
+ indirect, 3 * sizeof(GLuint));
+
+ brw->compute.num_work_groups_bo = bo;
+ brw->compute.num_work_groups_offset = indirect;
+ brw->compute.num_work_groups = indirect_group_counts;
+ ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
+
+ brw_dispatch_compute_common(ctx);
+}
void
brw_init_compute_functions(struct dd_function_table *functions)
{
functions->DispatchCompute = brw_dispatch_compute;
+ functions->DispatchComputeIndirect = brw_dispatch_compute_indirect;
}