radv: Support VK_EXT_inline_uniform_block.

Basically just reserve the memory in the descriptor sets.

On the shader side we construct a buffer descriptor, since
AFAIU VGPR indexing on 32-bit pointers in LLVM is still broken.

This fully supports update after bind and variable descriptor set
sizes. However, the limits are somewhat arbitrary and are mostly
about finding a reasonable division of a 2 GiB max memory size over
the set.

v2: - rebased on top of master (Samuel)
    - remove the loading resources rework (Samuel)
    - only load UBO descriptors if it's a pointer (Samuel)
    - use LLVMBuildPtrToInt to avoid IR failures (Samuel)

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> (v2)

5 files changed, 124 insertions, 15 deletions

diff --git a/src/amd/vulkan/radv_descriptor_set.c b/src/amd/vulkan/radv_descriptor_set.c
index 68171b5d244..6c6b88a4553 100644
--- a/src/amd/vulkan/radv_descriptor_set.c
+++ b/src/amd/vulkan/radv_descriptor_set.c
@@ -127,6 +127,7 @@ VkResult radv_CreateDescriptorSetLayout(
 		uint32_t b = binding->binding;
 		uint32_t alignment;
 		unsigned binding_buffer_count = 0;
+		uint32_t descriptor_count = binding->descriptorCount;
 
 		switch (binding->descriptorType) {
 		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
@@ -164,6 +165,11 @@ VkResult radv_CreateDescriptorSetLayout(
 			set_layout->binding[b].size = 16;
 			alignment = 16;
 			break;
+		case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
+			alignment = 16;
+			set_layout->binding[b].size = descriptor_count;
+			descriptor_count = 1;
+			break;
 		default:
 			unreachable("unknown descriptor type\n");
 			break;
@@ -171,7 +177,7 @@ VkResult radv_CreateDescriptorSetLayout(
 
 		set_layout->size = align(set_layout->size, alignment);
 		set_layout->binding[b].type = binding->descriptorType;
-		set_layout->binding[b].array_size = binding->descriptorCount;
+		set_layout->binding[b].array_size = descriptor_count;
 		set_layout->binding[b].offset = set_layout->size;
 		set_layout->binding[b].buffer_offset = buffer_count;
 		set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
@@ -207,9 +213,9 @@ VkResult radv_CreateDescriptorSetLayout(
 			samplers_offset += 4 * sizeof(uint32_t) * binding->descriptorCount;
 		}
 
-		set_layout->size += binding->descriptorCount * set_layout->binding[b].size;
-		buffer_count += binding->descriptorCount * binding_buffer_count;
-		dynamic_offset_count += binding->descriptorCount *
+		set_layout->size += descriptor_count * set_layout->binding[b].size;
+		buffer_count += descriptor_count * binding_buffer_count;
+		dynamic_offset_count += descriptor_count *
 			set_layout->binding[b].dynamic_offset_count;
 		set_layout->shader_stages |= binding->stageFlags;
 	}
@@ -264,6 +270,7 @@ void radv_GetDescriptorSetLayoutSupport(VkDevice device,
 
 		uint64_t descriptor_size = 0;
 		uint64_t descriptor_alignment = 1;
+		uint32_t descriptor_count = binding->descriptorCount;
 		switch (binding->descriptorType) {
 		case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 		case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
@@ -282,7 +289,7 @@ void radv_GetDescriptorSetLayoutSupport(VkDevice device,
 			descriptor_alignment = 32;
 			break;
 		case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-			if (!has_equal_immutable_samplers(binding->pImmutableSamplers, binding->descriptorCount)) {
+			if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) {
 				descriptor_size = 64;
 			} else {
 				descriptor_size = 96;
@@ -290,11 +297,16 @@ void radv_GetDescriptorSetLayoutSupport(VkDevice device,
 			descriptor_alignment = 32;
 			break;
 		case VK_DESCRIPTOR_TYPE_SAMPLER:
-			if (!has_equal_immutable_samplers(binding->pImmutableSamplers, binding->descriptorCount)) {
+			if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) {
 				descriptor_size = 16;
 				descriptor_alignment = 16;
 			}
 			break;
+		case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
+			descriptor_alignment = 16;
+			descriptor_size = descriptor_count;
+			descriptor_count = 1;
+			break;
 		default:
 			unreachable("unknown descriptor type\n");
 			break;
@@ -305,18 +317,20 @@ void radv_GetDescriptorSetLayoutSupport(VkDevice device,
 		}
 		size = align_u64(size, descriptor_alignment);
 
-		uint64_t max_count = UINT64_MAX;
-		if (descriptor_size)
-			max_count = (UINT64_MAX - size) / descriptor_size;
+		uint64_t max_count = INT32_MAX;
+		if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
+			max_count = INT32_MAX - size;
+		else if (descriptor_size)
+			max_count = (INT32_MAX - size) / descriptor_size;
 
-		if (max_count < binding->descriptorCount) {
+		if (max_count < descriptor_count) {
 			supported = false;
 		}
 		if (variable_flags && binding->binding <variable_flags->bindingCount && variable_count &&
 		    (variable_flags->pBindingFlags[binding->binding] & VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
 			variable_count->maxVariableDescriptorCount = MIN2(UINT32_MAX, max_count);
 		}
-		size += binding->descriptorCount * descriptor_size;
+		size += descriptor_count * descriptor_size;
 	}
 
 	free(bindings);
@@ -543,6 +557,21 @@ VkResult radv_CreateDescriptorPool(
 	uint64_t size = sizeof(struct radv_descriptor_pool);
 	uint64_t bo_size = 0, bo_count = 0, range_count = 0;
 
+	vk_foreach_struct(ext, pCreateInfo->pNext) {
+		switch (ext->sType) {
+		case VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_INLINE_UNIFORM_BLOCK_CREATE_INFO_EXT: {
+			const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT *info =
+				(const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT*)ext;
+			/* the sizes are 4 aligned, and we need to align to at
+			 * most 32, which needs at most 28 bytes extra per
+			 * binding. */
+			bo_size += 28llu * info->maxInlineUniformBlockBindings;
+			break;
+		}
+		default:
+			break;
+		}
+	}
 
 	for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
 		if (pCreateInfo->pPoolSizes[i].type != VK_DESCRIPTOR_TYPE_SAMPLER)
@@ -569,6 +598,9 @@ VkResult radv_CreateDescriptorPool(
 		case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 			bo_size += 96 * pCreateInfo->pPoolSizes[i].descriptorCount;
 			break;
+		case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
+			bo_size += pCreateInfo->pPoolSizes[i].descriptorCount;
+			break;
 		default:
 			unreachable("unknown descriptor type\n");
 			break;
@@ -764,6 +796,17 @@ static void write_buffer_descriptor(struct radv_device *device,
 		*buffer_list = buffer->bo;
 }
 
+static void write_block_descriptor(struct radv_device *device,
+                                   struct radv_cmd_buffer *cmd_buffer,
+                                   void *dst,
+                                   const VkWriteDescriptorSet *writeset)
+{
+	const VkWriteDescriptorSetInlineUniformBlockEXT *inline_ub =
+		vk_find_struct_const(writeset->pNext, WRITE_DESCRIPTOR_SET_INLINE_UNIFORM_BLOCK_EXT);
+
+	memcpy(dst, inline_ub->pData, inline_ub->dataSize);
+}
+
 static void write_dynamic_buffer_descriptor(struct radv_device *device,
                                             struct radv_descriptor_range *range,
                                             struct radeon_winsys_bo **buffer_list,
@@ -862,6 +905,12 @@ void radv_update_descriptor_sets(
 		const uint32_t *samplers = radv_immutable_samplers(set->layout, binding_layout);
 
 		ptr += binding_layout->offset / 4;
+
+		if (writeset->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+			write_block_descriptor(device, cmd_buffer, (uint8_t*)ptr + writeset->dstArrayElement, writeset);
+			continue;
+		}
+
 		ptr += binding_layout->size * writeset->dstArrayElement / 4;
 		buffer_list += binding_layout->buffer_offset;
 		buffer_list += writeset->dstArrayElement;
@@ -1042,7 +1091,12 @@ VkResult radv_CreateDescriptorUpdateTemplate(VkDevice _device,
 			default:
 				break;
 			}
-			dst_offset = binding_layout->offset / 4 + binding_layout->size * entry->dstArrayElement / 4;
+			dst_offset = binding_layout->offset / 4;
+			if (entry->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
+				dst_offset += entry->dstArrayElement / 4;
+			else
+				dst_offset += binding_layout->size * entry->dstArrayElement / 4;
+
 			dst_stride = binding_layout->size / 4;
 			break;
 		}
@@ -1092,6 +1146,11 @@ void radv_update_descriptor_set_with_template(struct radv_device *device,
 		const uint8_t *pSrc = ((const uint8_t *) pData) + templ->entry[i].src_offset;
 		uint32_t j;
 
+		if (templ->entry[i].descriptor_type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+			memcpy((uint8_t*)pDst, pSrc, templ->entry[i].descriptor_count);
+			continue;
+		}
+
 		for (j = 0; j < templ->entry[i].descriptor_count; ++j) {
 			switch (templ->entry[i].descriptor_type) {
 			case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index 96c6543141e..774ee5b91df 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -919,6 +919,14 @@ void radv_GetPhysicalDeviceFeatures2(
 			features->shaderSharedInt64Atomics = false;
 			break;
 		}
+		case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
+			VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
+				(VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
+
+			features->inlineUniformBlock = true;
+			features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
+			break;
+		}
 		default:
 			break;
 		}
@@ -1213,7 +1221,8 @@ void radv_GetPhysicalDeviceProperties2(
 			properties->robustBufferAccessUpdateAfterBind = false;
 			properties->quadDivergentImplicitLod = false;
 
-			size_t max_descriptor_set_size = ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
+			size_t max_descriptor_set_size = ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS -
+				MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
 			          (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
 			           32 /* storage buffer, 32 due to potential space wasted on alignment */ +
 			           32 /* sampler, largest when combined with image */ +
@@ -1301,6 +1310,17 @@ void radv_GetPhysicalDeviceProperties2(
 			properties->transformFeedbackDraw = true;
 			break;
 		}
+		case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
+			VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
+				(VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
+
+			props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
+			props->maxPerStageDescriptorInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_SETS;
+			props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_SETS;
+			props->maxDescriptorSetInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_COUNT;
+			props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_COUNT;
+			break;
+		}
 		default:
 			break;
 		}
diff --git a/src/amd/vulkan/radv_extensions.py b/src/amd/vulkan/radv_extensions.py
index 4b12ccc47a0..40fc585c503 100644
--- a/src/amd/vulkan/radv_extensions.py
+++ b/src/amd/vulkan/radv_extensions.py
@@ -112,6 +112,7 @@ EXTENSIONS = [
     Extension('VK_EXT_external_memory_host',              1, 'device->rad_info.has_userptr'),
     Extension('VK_EXT_global_priority',                   1, 'device->rad_info.has_ctx_priority'),
     Extension('VK_EXT_host_query_reset',                  1, True),
+    Extension('VK_EXT_inline_uniform_block',              1, True),
     Extension('VK_EXT_memory_budget',                     1, True),
     Extension('VK_EXT_memory_priority',                   1, True),
     Extension('VK_EXT_pci_bus_info',                      2, True),
diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
index 58a3cf18fe1..5bc88298ee6 100644
--- a/src/amd/vulkan/radv_nir_to_llvm.c
+++ b/src/amd/vulkan/radv_nir_to_llvm.c
@@ -1305,13 +1305,35 @@ radv_load_resource(struct ac_shader_abi *abi, LLVMValueRef index,
 	} else
 		stride = LLVMConstInt(ctx->ac.i32, layout->binding[binding].size, false);
 
-	offset = ac_build_imad(&ctx->ac, index, stride,
-			       LLVMConstInt(ctx->ac.i32, base_offset, false));
+	offset = LLVMConstInt(ctx->ac.i32, base_offset, false);
+
+	if (layout->binding[binding].type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+		offset = ac_build_imad(&ctx->ac, index, stride, offset);
+	}
 
 	desc_ptr = ac_build_gep0(&ctx->ac, desc_ptr, offset);
 	desc_ptr = ac_cast_ptr(&ctx->ac, desc_ptr, ctx->ac.v4i32);
 	LLVMSetMetadata(desc_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
 
+	if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+		uint32_t desc_type = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+			S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+			S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+			S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+			S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+			S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
+		LLVMValueRef desc_components[4] = {
+			LLVMBuildPtrToInt(ctx->ac.builder, desc_ptr, ctx->ac.intptr, ""),
+			LLVMConstInt(ctx->ac.i32, S_008F04_BASE_ADDRESS_HI(ctx->options->address32_hi), false),
+			/* High limit to support variable sizes. */
+			LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
+			LLVMConstInt(ctx->ac.i32, desc_type, false),
+		};
+
+		return ac_build_gather_values(&ctx->ac, desc_components, 4);
+	}
+
 	return desc_ptr;
 }
 
@@ -1910,6 +1932,11 @@ static LLVMValueRef radv_load_ubo(struct ac_shader_abi *abi, LLVMValueRef buffer
 	struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
 	LLVMValueRef result;
 
+	if (LLVMGetTypeKind(LLVMTypeOf(buffer_ptr)) != LLVMPointerTypeKind) {
+		/* Do not load the descriptor for inlined uniform blocks. */
+		return buffer_ptr;
+	}
+
 	LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
 
 	result = LLVMBuildLoad(ctx->ac.builder, buffer_ptr, "");
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index 82ab4eff8ca..cd3af7e614d 100644
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -102,6 +102,8 @@ typedef uint32_t xcb_window_t;
 #define MAX_SO_STREAMS 4
 #define MAX_SO_BUFFERS 4
 #define MAX_SO_OUTPUTS 64
+#define MAX_INLINE_UNIFORM_BLOCK_SIZE (4ull * 1024 * 1024)
+#define MAX_INLINE_UNIFORM_BLOCK_COUNT 64
 
 #define NUM_DEPTH_CLEAR_PIPELINES 3