radv: Expose VK_KHR_cooperative_matrix.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24683>

diff --git a/docs/relnotes/new_features.txt b/docs/relnotes/new_features.txt
index 62f5838..ca79cf2 100644 (file)
--- a/docs/relnotes/new_features.txt
+++ b/docs/relnotes/new_features.txt
@@ -18,3 +18,4 @@ GL_OES_sample_variables on Asahi
 GL_OES_shader_multisample_interpolation on Asahi
 GL_OES_gpu_shader5 on Asahi
 EGL_ANDROID_blob_cache works when disk caching is disabled
+VK_KHR_cooperative_matrix on RADV/GFX11+

diff --git a/src/amd/vulkan/radv_physical_device.c b/src/amd/vulkan/radv_physical_device.c
index ff897fd..3402415 100644 (file)
--- a/src/amd/vulkan/radv_physical_device.c
+++ b/src/amd/vulkan/radv_physical_device.c
@@ -380,6 +380,7 @@ radv_physical_device_get_supported_extensions(const struct radv_physical_device
       .KHR_8bit_storage = true,
       .KHR_16bit_storage = true,
       .KHR_acceleration_structure = radv_enable_rt(device, false),
+      .KHR_cooperative_matrix = device->rad_info.gfx_level >= GFX11 && !device->use_llvm,
       .KHR_bind_memory2 = true,
       .KHR_buffer_device_address = true,
       .KHR_copy_commands2 = true,
@@ -1038,6 +1039,10 @@ radv_physical_device_get_features(const struct radv_physical_device *pdevice, st
       .deviceGeneratedCompute = true,
       .deviceGeneratedComputePipelines = false,
       .deviceGeneratedComputeCaptureReplay = false,
+
+      /* VK_KHR_cooperative_matrix */
+      .cooperativeMatrix = pdevice->rad_info.gfx_level >= GFX11 && !pdevice->use_llvm,
+      .cooperativeMatrixRobustBufferAccess = pdevice->rad_info.gfx_level >= GFX11 && !pdevice->use_llvm,
    };
 }
 
@@ -1710,6 +1715,9 @@ radv_get_physical_device_properties(struct radv_physical_device *pdevice)
    p->polygonModePointSize = true;
    p->nonStrictSinglePixelWideLinesUseParallelogram = false;
    p->nonStrictWideLinesUseParallelogram = false;
+
+   /* VK_KHR_cooperative_matrix */
+   p->cooperativeMatrixSupportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
 }
 
 static VkResult
@@ -2436,3 +2444,65 @@ radv_GetPhysicalDeviceToolProperties(VkPhysicalDevice physicalDevice, uint32_t *
 
    return vk_outarray_status(&out);
 }
+
+VKAPI_ATTR VkResult VKAPI_CALL
+radv_GetPhysicalDeviceCooperativeMatrixPropertiesKHR(VkPhysicalDevice physicalDevice, uint32_t *pPropertyCount,
+                                                     VkCooperativeMatrixPropertiesKHR *pProperties)
+{
+   VK_OUTARRAY_MAKE_TYPED(VkCooperativeMatrixPropertiesKHR, out, pProperties, pPropertyCount);
+
+   vk_outarray_append_typed(VkCooperativeMatrixPropertiesKHR, &out, p)
+   {
+      *p = (struct VkCooperativeMatrixPropertiesKHR){.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR,
+                                                     .MSize = 16,
+                                                     .NSize = 16,
+                                                     .KSize = 16,
+                                                     .AType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .BType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .CType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .ResultType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .saturatingAccumulation = false,
+                                                     .scope = VK_SCOPE_SUBGROUP_KHR};
+   }
+
+   vk_outarray_append_typed(VkCooperativeMatrixPropertiesKHR, &out, p)
+   {
+      *p = (struct VkCooperativeMatrixPropertiesKHR){.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR,
+                                                     .MSize = 16,
+                                                     .NSize = 16,
+                                                     .KSize = 16,
+                                                     .AType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .BType = VK_COMPONENT_TYPE_FLOAT16_KHR,
+                                                     .CType = VK_COMPONENT_TYPE_FLOAT32_KHR,
+                                                     .ResultType = VK_COMPONENT_TYPE_FLOAT32_KHR,
+                                                     .saturatingAccumulation = false,
+                                                     .scope = VK_SCOPE_SUBGROUP_KHR};
+   }
+
+   for (unsigned asigned = 0; asigned < 2; asigned++) {
+      for (unsigned bsigned = 0; bsigned < 2; bsigned++) {
+         for (unsigned csigned = 0; csigned < 2; csigned++) {
+            for (unsigned saturate = 0; saturate < 2; saturate++) {
+               if (!csigned && saturate)
+                  continue; /* The HW only supports signed acc. */
+               vk_outarray_append_typed(VkCooperativeMatrixPropertiesKHR, &out, p)
+               {
+                  *p = (struct VkCooperativeMatrixPropertiesKHR){
+                     .sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR,
+                     .MSize = 16,
+                     .NSize = 16,
+                     .KSize = 16,
+                     .AType = asigned ? VK_COMPONENT_TYPE_SINT8_KHR : VK_COMPONENT_TYPE_UINT8_KHR,
+                     .BType = bsigned ? VK_COMPONENT_TYPE_SINT8_KHR : VK_COMPONENT_TYPE_UINT8_KHR,
+                     .CType = csigned ? VK_COMPONENT_TYPE_SINT32_KHR : VK_COMPONENT_TYPE_UINT32_KHR,
+                     .ResultType = csigned ? VK_COMPONENT_TYPE_SINT32_KHR : VK_COMPONENT_TYPE_UINT32_KHR,
+                     .saturatingAccumulation = saturate,
+                     .scope = VK_SCOPE_SUBGROUP_KHR};
+               }
+            }
+         }
+      }
+   }
+
+   return vk_outarray_status(&out);
+}