[Mesa-dev] [PATCH 6/6] radeonsi: fix is_oneway_access_only for bindless images

Marek Olšák maraeo at gmail.com
Tue Nov 20 04:09:42 UTC 2018


From: Marek Olšák <marek.olsak at amd.com>

---
 .../drivers/radeonsi/si_shader_tgsi_mem.c     | 29 +++++++++++++++----
 1 file changed, 23 insertions(+), 6 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index 979e47d3cc4..6decedc4cce 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -389,25 +389,29 @@ static void load_emit_memory(
 }
 
 /**
  * Return true if the memory accessed by a LOAD or STORE instruction is
  * read-only or write-only, respectively.
  *
  * \param shader_buffers_reverse_access_mask
  *	For LOAD, set this to (store | atomic) slot usage in the shader.
  *	For STORE, set this to (load | atomic) slot usage in the shader.
  * \param images_reverse_access_mask  Same as above, but for images.
+ * \param bindless_buffer_reverse_access_mask  Same as above, but for bindless image buffers.
+ * \param bindless_image_reverse_access_mask   Same as above, but for bindless images.
  */
 static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
 				  const struct tgsi_shader_info *info,
 				  unsigned shader_buffers_reverse_access_mask,
-				  unsigned images_reverse_access_mask)
+				  unsigned images_reverse_access_mask,
+				  bool bindless_buffer_reverse_access_mask,
+				  bool bindless_image_reverse_access_mask)
 {
 	enum tgsi_file_type resource_file;
 	unsigned resource_index;
 	bool resource_indirect;
 
 	if (inst->Instruction.Opcode == TGSI_OPCODE_STORE) {
 		resource_file = inst->Dst[0].Register.File;
 		resource_index = inst->Dst[0].Register.Index;
 		resource_indirect = inst->Dst[0].Register.Indirect;
 	} else {
@@ -421,25 +425,28 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
 	       /* bindless image */
 	       resource_file == TGSI_FILE_INPUT ||
 	       resource_file == TGSI_FILE_OUTPUT ||
 	       resource_file == TGSI_FILE_CONSTANT ||
 	       resource_file == TGSI_FILE_TEMPORARY ||
 	       resource_file == TGSI_FILE_IMMEDIATE);
 
 	assert(resource_file != TGSI_FILE_BUFFER ||
 	       inst->Memory.Texture == TGSI_TEXTURE_BUFFER);
 
+	bool bindless = resource_file != TGSI_FILE_BUFFER &&
+			resource_file != TGSI_FILE_IMAGE;
+
 	/* RESTRICT means NOALIAS.
 	 * If there are no writes, we can assume the accessed memory is read-only.
 	 * If there are no reads, we can assume the accessed memory is write-only.
 	 */
-	if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT) {
+	if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT && !bindless) {
 		unsigned reverse_access_mask;
 
 		if (resource_file == TGSI_FILE_BUFFER) {
 			reverse_access_mask = shader_buffers_reverse_access_mask;
 		} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
 			reverse_access_mask = info->images_buffers &
 					      images_reverse_access_mask;
 		} else {
 			reverse_access_mask = ~info->images_buffers &
 					      images_reverse_access_mask;
@@ -459,24 +466,26 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
 	 * buffers), it implies that buffer memory is read-only.
 	 * If there are no buffer reads (for both shader buffers & image
 	 * buffers), it implies that buffer memory is write-only.
 	 *
 	 * Same for the case when there are no writes/reads for non-buffer
 	 * images.
 	 */
 	if (resource_file == TGSI_FILE_BUFFER ||
 	    inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
 		if (!shader_buffers_reverse_access_mask &&
-		    !(info->images_buffers & images_reverse_access_mask))
+		    !(info->images_buffers & images_reverse_access_mask) &&
+		    !bindless_buffer_reverse_access_mask)
 			return true;
 	} else {
-		if (!(~info->images_buffers & images_reverse_access_mask))
+		if (!(~info->images_buffers & images_reverse_access_mask) &&
+		    !bindless_image_reverse_access_mask)
 			return true;
 	}
 	return false;
 }
 
 static void load_emit(
 		const struct lp_build_tgsi_action *action,
 		struct lp_build_tgsi_context *bld_base,
 		struct lp_build_emit_data *emit_data)
 {
@@ -515,21 +524,25 @@ static void load_emit(
 	}
 
 	if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
 		ac_build_waitcnt(&ctx->ac, VM_CNT);
 
 	can_speculate = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
 			  is_oneway_access_only(inst, info,
 						info->shader_buffers_store |
 						info->shader_buffers_atomic,
 						info->images_store |
-						info->images_atomic);
+						info->images_atomic,
+						info->uses_bindless_buffer_store |
+						info->uses_bindless_buffer_atomic,
+						info->uses_bindless_image_store |
+						info->uses_bindless_image_atomic);
 	args.cache_policy = get_cache_policy(ctx, inst, false, false, false);
 
 	if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
 		/* Don't use SMEM for shader buffer loads, because LLVM doesn't
 		 * select SMEM for SI.load.const with a non-constant offset, and
 		 * constant offsets practically don't exist with shader buffers.
 		 *
 		 * Also, SI.load.const doesn't use inst_offset when it's lowered
 		 * to VMEM, so we just end up with more VALU instructions in the end
 		 * and no benefit.
@@ -676,21 +689,25 @@ static void store_emit(
 
 	if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
 		store_emit_memory(ctx, emit_data);
 		return;
 	}
 
 	bool writeonly_memory = is_oneway_access_only(inst, info,
 						      info->shader_buffers_load |
 						      info->shader_buffers_atomic,
 						      info->images_load |
-						      info->images_atomic);
+						      info->images_atomic,
+						      info->uses_bindless_buffer_load |
+						      info->uses_bindless_buffer_atomic,
+						      info->uses_bindless_image_load |
+						      info->uses_bindless_image_atomic);
 	LLVMValueRef chans[4], value;
 	LLVMValueRef vindex = ctx->i32_0;
 	LLVMValueRef voffset = ctx->i32_0;
 	struct ac_image_args args = {};
 
 	for (unsigned chan = 0; chan < 4; ++chan)
 		chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);
 
 	value = ac_build_gather_values(&ctx->ac, chans, 4);
 
-- 
2.17.1



More information about the mesa-dev mailing list