AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

radv: move updating compute scratch for RT when stack size is emitted

Browse Source 
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31347>
This commit is contained in:
Samuel Pitoiset
2024-09-24 18:00:28 +02:00
committed by Marge Bot
parent ebe66dee08
commit f7482e85ba
1 changed files with 14 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/vulkan/radv_cmd_buffer.c
+14 -11
View File
@@ -11983,8 +11983,22 @@ static void
radv_emit_rt_stack_size(struct radv_cmd_buffer *cmd_buffer)
{
struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
const struct radv_physical_device *pdev = radv_device_physical(device);
const struct radv_shader *rt_prolog = cmd_buffer->state.rt_prolog;
unsigned rsrc2 = rt_prolog->config.rsrc2;
/* Reserve scratch for stacks manually since it is not handled by the compute path. */
uint32_t scratch_bytes_per_wave = rt_prolog->config.scratch_bytes_per_wave;
const uint32_t wave_size = rt_prolog->info.wave_size;
/* The hardware register is specified as a multiple of 64 or 256 DWORDS. */
const unsigned scratch_alloc_granule = pdev->info.gfx_level >= GFX11 ? 256 : 1024;
scratch_bytes_per_wave += align(cmd_buffer->state.rt_stack_size * wave_size, scratch_alloc_granule);
cmd_buffer->compute_scratch_size_per_wave_needed =
MAX2(cmd_buffer->compute_scratch_size_per_wave_needed, scratch_bytes_per_wave);
if (cmd_buffer->state.rt_stack_size)
rsrc2 |= S_00B12C_SCRATCH_EN(1);
@@ -12287,17 +12301,6 @@ radv_trace_rays(struct radv_cmd_buffer *cmd_buffer, VkTraceRaysIndirectCommand2K
struct radv_compute_pipeline *pipeline = &cmd_buffer->state.rt_pipeline->base;
struct radv_shader *rt_prolog = cmd_buffer->state.rt_prolog;
/* Reserve scratch for stacks manually since it is not handled by the compute path. */
uint32_t scratch_bytes_per_wave = rt_prolog->config.scratch_bytes_per_wave;
uint32_t wave_size = rt_prolog->info.wave_size;
/* The hardware register is specified as a multiple of 64 or 256 DWORDS. */
unsigned scratch_alloc_granule = pdev->info.gfx_level >= GFX11 ? 256 : 1024;
scratch_bytes_per_wave += align(cmd_buffer->state.rt_stack_size * wave_size, scratch_alloc_granule);
cmd_buffer->compute_scratch_size_per_wave_needed =
MAX2(cmd_buffer->compute_scratch_size_per_wave_needed, scratch_bytes_per_wave);
/* Since the workgroup size is 8x4 (or 8x8), 1D dispatches can only fill 8 threads per wave at most. To increase
* occupancy, it's beneficial to convert to a 2D dispatch in these cases. */
if (tables && tables->height == 1 && tables->width >= cmd_buffer->state.rt_prolog->info.cs.block_size[0])