radv: use a more relaxed alignment for upload buffer allocations

256 bytes was higher than necessary.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8833>

src/amd/vulkan/radv_cmd_buffer.c

@@ -459,7 +459,7 @@ radv_reset_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
 		unsigned fence_offset, eop_bug_offset;
 		void *fence_ptr;
 
-		radv_cmd_buffer_upload_alloc(cmd_buffer, 8, 8, &fence_offset,
+		radv_cmd_buffer_upload_alloc(cmd_buffer, 8, &fence_offset,
 					     &fence_ptr);
 		memset(fence_ptr, 0, 8);
 
@@ -469,7 +469,7 @@ radv_reset_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
 
 		if (cmd_buffer->device->physical_device->rad_info.chip_class == GFX9) {
 			/* Allocate a buffer for the EOP bug on GFX9. */
-			radv_cmd_buffer_upload_alloc(cmd_buffer, 16 * num_db, 8,
+			radv_cmd_buffer_upload_alloc(cmd_buffer, 16 * num_db,
 						     &eop_bug_offset, &fence_ptr);
 			memset(fence_ptr, 0, 16 * num_db);
 			cmd_buffer->gfx9_eop_bug_va =
@@ -548,14 +548,21 @@ radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer *cmd_buffer,
 
 bool
 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
-			     unsigned size,
+			     unsigned size, unsigned *out_offset, void **ptr)
-			     unsigned alignment,
-			     unsigned *out_offset,
-			     void **ptr)
 {
-	assert(util_is_power_of_two_nonzero(alignment));
+	assert(size % 4 == 0);
+
+	struct radeon_info *rad_info = &cmd_buffer->device->physical_device->rad_info;
+
+	/* Align to the scalar cache line size if it results in this allocation
+	 * being placed in less of them.
+	 */
+	unsigned offset = cmd_buffer->upload.offset;
+	unsigned line_size = rad_info->chip_class >= GFX10 ? 64 : 32;
+	unsigned gap = align(offset, line_size) - offset;
+	if ((size & (line_size - 1)) > gap)
+		offset = align(offset, line_size);
 
-	uint64_t offset = align(cmd_buffer->upload.offset, alignment);
 	if (offset + size > cmd_buffer->upload.size) {
 		if (!radv_cmd_buffer_resize_upload_buf(cmd_buffer, size))
 			return false;
@@ -571,13 +578,11 @@ radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
 
 bool
 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
-			    unsigned size, unsigned alignment,
+			    unsigned size, const void *data, unsigned *out_offset)
-			    const void *data, unsigned *out_offset)
 {
 	uint8_t *ptr;
 
-	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, alignment,
+	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, out_offset, (void **)&ptr))
-					  out_offset, (void **)&ptr))
 		return false;
 
 	if (ptr)
@@ -2634,7 +2639,7 @@ radv_flush_push_descriptors(struct radv_cmd_buffer *cmd_buffer,
 		(struct radv_descriptor_set *)&descriptors_state->push_set.set;
 	unsigned bo_offset;
 
-	if (!radv_cmd_buffer_upload_data(cmd_buffer, set->header.size, 32,
+	if (!radv_cmd_buffer_upload_data(cmd_buffer, set->header.size,
 					 set->header.mapped_ptr,
 					 &bo_offset))
 		return;
@@ -2653,8 +2658,7 @@ radv_flush_indirect_descriptor_sets(struct radv_cmd_buffer *cmd_buffer,
 	uint32_t offset;
 	void *ptr;
 
-	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size,
+	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, &offset, &ptr))
-					  256, &offset, &ptr))
 		return;
 
 	for (unsigned i = 0; i < MAX_SETS; i++) {
@@ -2798,8 +2802,7 @@ radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
 
 	if (need_push_constants) {
 		if (!radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
-						  16 * layout->dynamic_offset_count,
+						  16 * layout->dynamic_offset_count, &offset, &ptr))
-						  256, &offset, &ptr))
 			return;
 
 		memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
@@ -2847,7 +2850,7 @@ radv_flush_vertex_descriptors(struct radv_cmd_buffer *cmd_buffer,
 		uint64_t va;
 
 		/* allocate some descriptor state for vertex buffers */
-		if (!radv_cmd_buffer_upload_alloc(cmd_buffer, count * 16, 256,
+		if (!radv_cmd_buffer_upload_alloc(cmd_buffer, count * 16,
 						  &vb_offset, &vb_ptr))
 			return;
 
@@ -2970,7 +2973,7 @@ radv_flush_streamout_descriptors(struct radv_cmd_buffer *cmd_buffer)
 
 		/* Allocate some descriptor state for streamout buffers. */
 		if (!radv_cmd_buffer_upload_alloc(cmd_buffer,
-						  MAX_SO_BUFFERS * 16, 256,
+						  MAX_SO_BUFFERS * 16,
 						  &so_offset, &so_ptr))
 			return;
 
@@ -4237,9 +4240,8 @@ void radv_meta_push_descriptor_set(
 	push_set->header.size = layout->set[set].layout->size;
 	push_set->header.layout = layout->set[set].layout;
 
-	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, push_set->header.size, 32,
+	if (!radv_cmd_buffer_upload_alloc(cmd_buffer, push_set->header.size,
-	                                  &bo_offset,
+	                                  &bo_offset, (void**) &push_set->header.mapped_ptr))
-	                                  (void**) &push_set->header.mapped_ptr))
 		return;
 
 	push_set->header.va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
@@ -6789,7 +6791,7 @@ void radv_CmdBeginConditionalRenderingEXT(
 		 * Based on the conditionalrender demo, it's faster to do the
 		 * COPY_DATA in ME  (+ sync PFP) instead of PFP.
 		 */
-		radv_cmd_buffer_upload_data(cmd_buffer, 8, 16, &pred_value, &pred_offset);
+		radv_cmd_buffer_upload_data(cmd_buffer, 8, &pred_value, &pred_offset);
 
 		pred_va = radv_buffer_get_va(cmd_buffer->upload.upload_bo) + pred_offset;
 

src/amd/vulkan/radv_meta_buffer.c

@@ -546,7 +546,7 @@ void radv_CmdUpdateBuffer(
 			radv_cmd_buffer_trace_emit(cmd_buffer);
 	} else {
 		uint32_t buf_offset;
-		radv_cmd_buffer_upload_data(cmd_buffer, dataSize, 32, pData, &buf_offset);
+		radv_cmd_buffer_upload_data(cmd_buffer, dataSize, pData, &buf_offset);
 		radv_copy_buffer(cmd_buffer, cmd_buffer->upload.upload_bo, dst_buffer->bo,
 				 buf_offset, dstOffset + dst_buffer->offset, dataSize);
 	}

src/amd/vulkan/radv_private.h

@@ -1504,17 +1504,13 @@ void si_cp_dma_wait_for_idle(struct radv_cmd_buffer *cmd_buffer);
 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
 bool
 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
-			     unsigned size,
+			     unsigned size, unsigned *out_offset, void **ptr);
-			     unsigned alignment,
-			     unsigned *out_offset,
-			     void **ptr);
 void
 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
 			    const struct radv_subpass *subpass);
 bool
 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
-			    unsigned size, unsigned alignmnet,
+			    unsigned size, const void *data, unsigned *out_offset);
-			    const void *data, unsigned *out_offset);
 
 void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
 void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);

src/amd/vulkan/si_cmd_buffer.c

@@ -1744,7 +1744,7 @@ static void si_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigne
 
 	assert(size < SI_CPDMA_ALIGNMENT);
 
-	radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, SI_CPDMA_ALIGNMENT,  &offset, &ptr);
+	radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size,  &offset, &ptr);
 
 	va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
 	va += offset;