fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-21 19:30:38 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 5

anv: Implement VK_KHR_cooperative_matrix

Browse source 
v2: Rebase on moving lowering pass to src/intel/compiler.

v3: Don't hide the extension behind an environment variable
(ANV_COOPERATIVE_MATRIX) now the we have a better solution for setting
PIPELINE_SELECT.

v4: Prefix type names with INTEL_CMAT_. Suggested by Lionel. Also rebase
on f99e43d606 ("anv: switch to use runtime physical device properties
infrastructure").

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25994>
This commit is contained in:
Caio Oliveira 2023-06-13 19:48:16 -07:00 • committed by Ian Romanick
parent ff16458478
commit 0a6f8b40bf
3 changed files with 88 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

75
src/intel/vulkan/anv_device.c
View file

@ -218,6 +218,9 @@ get_device_extensions(const struct anv_physical_device *device,
const bool rt_enabled = ANV_SUPPORT_RT && device->info.has_ray_tracing;
const bool cooperative_matrix_enabled =
anv_has_cooperative_matrix(&device->info);
*ext = (struct vk_device_extension_table) {
.KHR_8bit_storage = true,
.KHR_16bit_storage = !device->instance->no_16bit,
@ -226,6 +229,7 @@ get_device_extensions(const struct anv_physical_device *device,
.KHR_buffer_device_address = true,
.KHR_calibrated_timestamps = device->has_reg_timestamp,
.KHR_copy_commands2 = true,
.KHR_cooperative_matrix = cooperative_matrix_enabled,
.KHR_create_renderpass2 = true,
.KHR_dedicated_allocation = true,
.KHR_deferred_host_operations = true,
@ -862,6 +866,9 @@ get_features(const struct anv_physical_device *pdevice,
.nestedCommandBuffer = true,
.nestedCommandBufferRendering = true,
.nestedCommandBufferSimultaneousUse = false,
/* VK_KHR_cooperative_matrix */
.cooperativeMatrix = anv_has_cooperative_matrix(&pdevice->info),
};
/* The new DOOM and Wolfenstein games require depthBounds without
@ -1319,6 +1326,9 @@ get_properties(const struct anv_physical_device *pdevice,
.sparseResidencyStandard3DBlockShape = has_sparse_or_fake,
.sparseResidencyAlignedMipSize = false,
.sparseResidencyNonResidentStrict = has_sparse_or_fake,
/* VK_KHR_cooperative_matrix */
.cooperativeMatrixSupportedStages = VK_SHADER_STAGE_COMPUTE_BIT,
};
snprintf(props->deviceName, sizeof(props->deviceName),
@ -5127,3 +5137,68 @@ anv_device_get_pat_entry(struct anv_device *device,
else
return &device->info->pat.writecombining;
}
static VkComponentTypeKHR
convert_component_type(enum intel_cooperative_matrix_component_type t)
{
switch (t) {
case INTEL_CMAT_FLOAT16: return VK_COMPONENT_TYPE_FLOAT16_KHR;
case INTEL_CMAT_FLOAT32: return VK_COMPONENT_TYPE_FLOAT32_KHR;
case INTEL_CMAT_SINT32: return VK_COMPONENT_TYPE_SINT32_KHR;
case INTEL_CMAT_SINT8: return VK_COMPONENT_TYPE_SINT8_KHR;
case INTEL_CMAT_UINT32: return VK_COMPONENT_TYPE_UINT32_KHR;
case INTEL_CMAT_UINT8: return VK_COMPONENT_TYPE_UINT8_KHR;
}
unreachable("invalid cooperative matrix component type in configuration");
}
static VkScopeKHR
convert_scope(mesa_scope scope)
{
switch (scope) {
case SCOPE_DEVICE: return VK_SCOPE_DEVICE_KHR;
case SCOPE_WORKGROUP: return VK_SCOPE_WORKGROUP_KHR;
case SCOPE_SUBGROUP: return VK_SCOPE_SUBGROUP_KHR;
case SCOPE_QUEUE_FAMILY: return VK_SCOPE_QUEUE_FAMILY_KHR;
default:
unreachable("invalid cooperative matrix scope in configuration");
}
}
VkResult anv_GetPhysicalDeviceCooperativeMatrixPropertiesKHR(
VkPhysicalDevice physicalDevice,
uint32_t* pPropertyCount,
VkCooperativeMatrixPropertiesKHR* pProperties)
{
ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
const struct intel_device_info *devinfo = &pdevice->info;
assert(anv_has_cooperative_matrix(devinfo));
VK_OUTARRAY_MAKE_TYPED(VkCooperativeMatrixPropertiesKHR, out, pProperties, pPropertyCount);
for (int i = 0; i < ARRAY_SIZE(devinfo->cooperative_matrix_configurations); i++) {
const struct intel_cooperative_matrix_configuration *cfg =
&devinfo->cooperative_matrix_configurations[i];
if (cfg->scope == SCOPE_NONE)
break;
vk_outarray_append_typed(VkCooperativeMatrixPropertiesKHR, &out, prop) {
prop->sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR;
prop->MSize = cfg->m;
prop->NSize = cfg->n;
prop->KSize = cfg->k;
prop->AType = convert_component_type(cfg->a);
prop->BType = convert_component_type(cfg->b);
prop->CType = convert_component_type(cfg->c);
prop->ResultType = convert_component_type(cfg->result);
prop->scope = convert_scope(cfg->scope);
}
}
return vk_outarray_status(&out);
}

8
src/intel/vulkan/anv_pipeline.c
View file

@ -144,6 +144,7 @@ anv_shader_stage_to_nir(struct anv_device *device,
const bool rt_enabled = ANV_SUPPORT_RT && pdevice->info.has_ray_tracing;
const struct spirv_to_nir_options spirv_options = {
.caps = {
.cooperative_matrix = anv_has_cooperative_matrix(&pdevice->info),
.demote_to_helper_invocation = true,
.derivative_group = true,
.descriptor_array_dynamic_indexing = true,
@ -977,6 +978,13 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
use_primitive_replication);
}
if (gl_shader_stage_uses_workgroup(nir->info.stage)) {
// TODO(coop): Write this restriction the right way, maybe we need to do at the backend.
assert(nir->info.subgroup_size <= SUBGROUP_SIZE_REQUIRE_8);
nir->info.subgroup_size = SUBGROUP_SIZE_REQUIRE_8;
NIR_PASS(_, nir, brw_nir_lower_cmat, nir->info.subgroup_size);
}
/* The patch control points are delivered through a push constant when
* dynamic.
*/

5
src/intel/vulkan/anv_private.h
View file

@ -5804,6 +5804,11 @@ static inline void anv_perfetto_end_submit(struct anv_queue *queue,
{}
#endif
static bool
anv_has_cooperative_matrix(const struct intel_device_info *info)
{
return info->cooperative_matrix_configurations[0].scope != SCOPE_NONE;
}
#define ANV_FROM_HANDLE(__anv_type, __name, __handle) \
VK_FROM_HANDLE(__anv_type, __name, __handle)