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

iris, i965: Update limits for ARB_compute_variable_group_size

Browse Source 
The CS compiler now produces multiple SIMD variants, so the previous
trade-off between "always using SIMD32" and "having a smaller max
invocations" is now gone.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5142>
This commit is contained in:
Caio Marcelo de Oliveira Filho
2020-05-20 23:32:18 -07:00
parent 46b428074f
commit 78e400d4a5
3 changed files with 11 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/iris/iris_screen.c
+5 -31
View File
@@ -435,32 +435,6 @@ iris_get_shader_param(struct pipe_screen *pscreen,
}
}
static unsigned
get_max_threads(const struct gen_device_info *devinfo)
{
/* Limit max_threads to 64 for the GPGPU_WALKER command. */
return MIN2(64, devinfo->max_cs_threads);
}
uint32_t
iris_get_max_var_invocations(const struct iris_screen *screen)
{
const unsigned max_threads = get_max_threads(&screen->devinfo);
/* 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.
*/
const uint32_t max_var_invocations =
(max_threads >= 64 ? 8 : (max_threads >= 32 ? 16 : 32)) * max_threads;
assert(max_var_invocations >= 512);
return max_var_invocations;
}
static int
iris_get_compute_param(struct pipe_screen *pscreen,
enum pipe_shader_ir ir_type,
@@ -468,8 +442,10 @@ iris_get_compute_param(struct pipe_screen *pscreen,
void *ret)
{
struct iris_screen *screen = (struct iris_screen *)pscreen;
const struct gen_device_info *devinfo = &screen->devinfo;
const unsigned max_threads = get_max_threads(&screen->devinfo);
/* Limit max_threads to 64 for the GPGPU_WALKER command. */
const unsigned max_threads = MIN2(64, devinfo->max_cs_threads);
const uint32_t max_invocations = 32 * max_threads;
#define RET(x) do { \
@@ -499,6 +475,8 @@ iris_get_compute_param(struct pipe_screen *pscreen,
case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
/* MaxComputeWorkGroupInvocations */
case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
/* MaxComputeVariableGroupInvocations */
RET((uint64_t []) { max_invocations });
case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
@@ -511,10 +489,6 @@ iris_get_compute_param(struct pipe_screen *pscreen,
case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
RET((uint32_t []) { BRW_SUBGROUP_SIZE });
case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
/* MaxComputeVariableGroupInvocations */
RET((uint64_t []) { iris_get_max_var_invocations(screen) });
case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
src/gallium/drivers/iris/iris_screen.h
-2
View File
@@ -238,6 +238,4 @@ iris_is_format_supported(struct pipe_screen *pscreen,
void iris_disk_cache_init(struct iris_screen *screen);
uint32_t iris_get_max_var_invocations(const struct iris_screen *screen);
#endif
src/mesa/drivers/dri/i965/brw_context.c
+6 -15
View File
@@ -836,22 +836,13 @@ brw_initialize_cs_context_constants(struct brw_context *brw)
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.
*/
/* Constants used for ARB_compute_variable_group_size. */
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;
assert(max_invocations >= 512);
ctx->Const.MaxComputeVariableGroupSize[0] = max_invocations;
ctx->Const.MaxComputeVariableGroupSize[1] = max_invocations;
ctx->Const.MaxComputeVariableGroupSize[2] = max_invocations;
ctx->Const.MaxComputeVariableGroupInvocations = max_invocations;
}
}