fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-06 11:38:05 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

v3d,v3dv: fix compute for V3D 7.1.6+

Browse source 
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25450>
This commit is contained in:
Iago Toral Quiroga 2023-03-09 19:05:19 +01:00 • committed by Marge Bot
parent 8a69888803
commit d25aec0d38
4 changed files with 35 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
src/broadcom/vulkan/v3dv_cmd_buffer.c
View file

@ -4124,6 +4124,7 @@ cmd_buffer_emit_pre_dispatch(struct v3dv_cmd_buffer *cmd_buffer)
void
v3dv_cmd_buffer_rewrite_indirect_csd_job(
struct v3dv_device *device,
struct v3dv_csd_indirect_cpu_job_info *info,
const uint32_t *wg_counts)
{
@ -4143,8 +4144,15 @@ v3dv_cmd_buffer_rewrite_indirect_csd_job(
submit->cfg[1] = wg_counts[1] << V3D_CSD_CFG012_WG_COUNT_SHIFT;
submit->cfg[2] = wg_counts[2] << V3D_CSD_CFG012_WG_COUNT_SHIFT;
submit->cfg[4] = DIV_ROUND_UP(info->wg_size, 16) *
(wg_counts[0] * wg_counts[1] * wg_counts[2]) - 1;
uint32_t num_batches = DIV_ROUND_UP(info->wg_size, 16) *
(wg_counts[0] * wg_counts[1] * wg_counts[2]);
/* V3D 7.1.6 and later don't subtract 1 from the number of batches */
if (device->devinfo.ver < 71 ||
(device->devinfo.ver == 71 && device->devinfo.rev < 6)) {
submit->cfg[4] = num_batches - 1;
} else {
submit->cfg[4] = num_batches;
}
assert(submit->cfg[4] != ~0);
if (info->needs_wg_uniform_rewrite) {
@ -4177,6 +4185,7 @@ cmd_buffer_create_csd_job(struct v3dv_cmd_buffer *cmd_buffer,
uint32_t **wg_uniform_offsets_out,
uint32_t *wg_size_out)
{
struct v3dv_device *device = cmd_buffer->device;
struct v3dv_pipeline *pipeline = cmd_buffer->state.compute.pipeline;
assert(pipeline && pipeline->shared_data->variants[BROADCOM_SHADER_COMPUTE]);
struct v3dv_shader_variant *cs_variant =
@ -4235,18 +4244,26 @@ cmd_buffer_create_csd_job(struct v3dv_cmd_buffer *cmd_buffer,
if (wg_size_out)
*wg_size_out = wg_size;
submit->cfg[4] = num_batches - 1;
/* V3D 7.1.6 and later don't subtract 1 from the number of batches */
if (device->devinfo.ver < 71 ||
(device->devinfo.ver == 71 && device->devinfo.rev < 6)) {
submit->cfg[4] = num_batches - 1;
} else {
submit->cfg[4] = num_batches;
}
assert(submit->cfg[4] != ~0);
assert(pipeline->shared_data->assembly_bo);
struct v3dv_bo *cs_assembly_bo = pipeline->shared_data->assembly_bo;
submit->cfg[5] = cs_assembly_bo->offset + cs_variant->assembly_offset;
submit->cfg[5] |= V3D_CSD_CFG5_PROPAGATE_NANS;
if (cs_variant->prog_data.base->single_seg)
submit->cfg[5] |= V3D_CSD_CFG5_SINGLE_SEG;
if (cs_variant->prog_data.base->threads == 4)
submit->cfg[5] |= V3D_CSD_CFG5_THREADING;
/* V3D 7.x has made the PROPAGATE_NANS bit in CFG5 reserved */
if (device->devinfo.ver < 71)
submit->cfg[5] |= V3D_CSD_CFG5_PROPAGATE_NANS;
if (cs_variant->prog_data.cs->shared_size > 0) {
job->csd.shared_memory =

3
src/broadcom/vulkan/v3dv_private.h
View file

@ -1816,7 +1816,8 @@ void v3dv_cmd_buffer_copy_query_results(struct v3dv_cmd_buffer *cmd_buffer,
void v3dv_cmd_buffer_add_tfu_job(struct v3dv_cmd_buffer *cmd_buffer,
struct drm_v3d_submit_tfu *tfu);
void v3dv_cmd_buffer_rewrite_indirect_csd_job(struct v3dv_csd_indirect_cpu_job_info *info,
void v3dv_cmd_buffer_rewrite_indirect_csd_job(struct v3dv_device *device,
struct v3dv_csd_indirect_cpu_job_info *info,
const uint32_t *wg_counts);
void v3dv_cmd_buffer_add_private_obj(struct v3dv_cmd_buffer *cmd_buffer,

2
src/broadcom/vulkan/v3dv_queue.c
View file

@ -354,7 +354,7 @@ handle_csd_indirect_cpu_job(struct v3dv_queue *queue,
if (memcmp(group_counts, info->csd_job->csd.wg_count,
sizeof(info->csd_job->csd.wg_count)) != 0) {
v3dv_cmd_buffer_rewrite_indirect_csd_job(info, group_counts);
v3dv_cmd_buffer_rewrite_indirect_csd_job(queue->device, info, group_counts);
}
return VK_SUCCESS;

14
src/gallium/drivers/v3d/v3dx_draw.c
View file

@ -1473,8 +1473,15 @@ v3d_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
submit.cfg[3] |= (wg_size & 0xff) << V3D_CSD_CFG3_WG_SIZE_SHIFT;
/* Number of batches the dispatch will invoke (minus 1). */
submit.cfg[4] = num_batches - 1;
/* Number of batches the dispatch will invoke.
* V3D 7.1.6 and later don't subtract 1 from the number of batches
*/
if (v3d->screen->devinfo.ver < 71 ||
(v3d->screen->devinfo.ver == 71 && v3d->screen->devinfo.rev < 6)) {
submit.cfg[4] = num_batches - 1;
} else {
submit.cfg[4] = num_batches;
}
/* Make sure we didn't accidentally underflow. */
assert(submit.cfg[4] != ~0);
@ -1482,7 +1489,8 @@ v3d_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
v3d_job_add_bo(job, v3d_resource(v3d->prog.compute->resource)->bo);
submit.cfg[5] = (v3d_resource(v3d->prog.compute->resource)->bo->offset +
v3d->prog.compute->offset);
submit.cfg[5] |= V3D_CSD_CFG5_PROPAGATE_NANS;
if (v3d->screen->devinfo.ver < 71)
submit.cfg[5] |= V3D_CSD_CFG5_PROPAGATE_NANS;
if (v3d->prog.compute->prog_data.base->single_seg)
submit.cfg[5] |= V3D_CSD_CFG5_SINGLE_SEG;
if (v3d->prog.compute->prog_data.base->threads == 4)