gallivm: Optimize single-invocation SSBO stores.

The CTS does a lot of 1x1x1 compute shaders (all that stuff like dEQP-GLES31.functional.shaders.builtin_functions.precision.mul.highp_compute.scalar) which finish with store_ssbos. Instead of doing the invocation loop in that case (which LLVM has to later unroll), just emit the single invocation's store. Fixes timeouts running dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36, which does a spectacular number of SSBO stores in a long 1x1x1 compute shader. Reduces runtime of on llvmpipe from 66s to 29s locally, and virgl from 1:38 to 43s. virgl dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22 goes down to 7 seconds. Fixes: #6797 Reviewed-by: Dave Airlie <airlied@redhat.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17730>
2022-07-24 08:11:49 -07:00 · 2022-07-24 08:11:49 -07:00 · bd8740da77
parent 8d41f8f384
commit bd8740da77
9 changed files with 52 additions and 10 deletions
--- a/src/gallium/auxiliary/gallivm/lp_bld_nir.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_nir.c
@ -1550,10 +1550,11 @@ visit_store_ssbo(struct lp_build_nir_context *bld_base,
   LLVMValueRef val = get_src(bld_base, instr->src[0]);
   LLVMValueRef idx = cast_type(bld_base, get_src(bld_base, instr->src[1]), nir_type_uint, 32);
   LLVMValueRef offset = get_src(bld_base, instr->src[2]);
+   bool index_and_offset_are_uniform = nir_src_is_always_uniform(instr->src[1]) && nir_src_is_always_uniform(instr->src[2]);
   int writemask = instr->const_index[0];
   int nc = nir_src_num_components(instr->src[0]);
   int bitsize = nir_src_bit_size(instr->src[0]);
-   bld_base->store_mem(bld_base, writemask, nc, bitsize, idx, offset, val);
+   bld_base->store_mem(bld_base, writemask, nc, bitsize, index_and_offset_are_uniform, idx, offset, val);
 }


@ -1821,10 +1822,11 @@ visit_shared_store(struct lp_build_nir_context *bld_base,
 {
   LLVMValueRef val = get_src(bld_base, instr->src[0]);
   LLVMValueRef offset = get_src(bld_base, instr->src[1]);
+   bool offset_is_uniform = nir_src_is_always_uniform(instr->src[1]);
   int writemask = instr->const_index[1];
   int nc = nir_src_num_components(instr->src[0]);
   int bitsize = nir_src_bit_size(instr->src[0]);
-   bld_base->store_mem(bld_base, writemask, nc, bitsize, NULL, offset, val);
+   bld_base->store_mem(bld_base, writemask, nc, bitsize, offset_is_uniform, NULL, offset, val);
 }


--- a/src/gallium/auxiliary/gallivm/lp_bld_nir.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_nir.h
@ -115,6 +115,7 @@ struct lp_build_nir_context
                    LLVMValueRef index, LLVMValueRef offset, LLVMValueRef result[NIR_MAX_VEC_COMPONENTS]);
   void (*store_mem)(struct lp_build_nir_context *bld_base,
                     unsigned writemask, unsigned nc, unsigned bit_size,
+                     bool index_and_offset_are_uniform,
                     LLVMValueRef index, LLVMValueRef offset, LLVMValueRef dst);

   void (*atomic_mem)(struct lp_build_nir_context *bld_base,
--- a/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c
@ -1296,6 +1296,7 @@ static void emit_store_mem(struct lp_build_nir_context *bld_base,
                           unsigned writemask,
                           unsigned nc,
                           unsigned bit_size,
+                           bool index_and_offset_are_uniform,
                           LLVMValueRef index,
                           LLVMValueRef offset,
                           LLVMValueRef dst)
@ -1310,6 +1311,41 @@ static void emit_store_mem(struct lp_build_nir_context *bld_base,

   offset = lp_build_shr_imm(uint_bld, offset, shift_val);

+   /* If the address is uniform, then just store the value from the first
+    * channel instead of making LLVM unroll the invocation loop.
+    */
+   if (index_and_offset_are_uniform && invocation_0_must_be_active(bld_base)) {
+      LLVMValueRef ssbo_limit;
+      LLVMValueRef mem_ptr = mem_access_base_pointer(bld_base, store_bld, bit_size, index,
+                                                     lp_build_const_int32(gallivm, 0), &ssbo_limit);
+
+      offset = LLVMBuildExtractElement(gallivm->builder, offset, lp_build_const_int32(gallivm, 0), "");
+
+      for (unsigned c = 0; c < nc; c++) {
+         if (!(writemask & (1u << c)))
+            continue;
+
+         /* Pick out invocation 0's value. */
+         LLVMValueRef val = (nc == 1) ? dst : LLVMBuildExtractValue(builder, dst, c, "");
+         LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
+                                                         lp_build_const_int32(gallivm, 0), "");
+         value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, store_bld->elem_type, "");
+
+         LLVMValueRef chan_offset = LLVMBuildAdd(builder, offset, lp_build_const_int32(gallivm, c), "");
+
+         /* If storing outside the SSBO, we need to skip the store instead. */
+         if (ssbo_limit) {
+            struct lp_build_if_state ifthen;
+            lp_build_if(&ifthen, gallivm, lp_offset_in_range(bld_base, chan_offset, ssbo_limit));
+            lp_build_pointer_set(builder, mem_ptr, chan_offset, value_ptr);
+            lp_build_endif(&ifthen);
+         } else {
+            lp_build_pointer_set(builder, mem_ptr, chan_offset, value_ptr);
+         }
+      }
+      return;
+   }
+
   LLVMValueRef exec_mask = mask_vec(bld_base);
   LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
   struct lp_build_loop_state loop_state;
--- a/src/gallium/drivers/llvmpipe/ci/llvmpipe-skips.txt
+++ b/src/gallium/drivers/llvmpipe/ci/llvmpipe-skips.txt
@ -13,7 +13,6 @@ KHR-GL45.texture_size_promotion.functional
 KHR-GL45.texture_swizzle.functional
 KHR-GL45.texture_swizzle.smoke
 KHR-GLES31.core.arrays_of_arrays.InteractionFunctionCalls2
-dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36
 arb_pipeline_statistics_query-comp
 gl-1.0-blend-func

--- a/src/gallium/drivers/virgl/ci/virgl-gl-fails.txt
+++ b/src/gallium/drivers/virgl/ci/virgl-gl-fails.txt
@ -36,7 +36,10 @@ dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_rbo_2,Fail
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_texture_1,Fail
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_texture_2,Fail
+
+# Times out waiting for >15s compile on the host side.
 dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36,Fail
+dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22,Fail

 KHR-GL30.transform_feedback.api_errors_test,Fail
 KHR-GL32.transform_feedback_overflow_query_ARB.advanced-single-stream-interleaved-attribs,Fail
--- a/src/gallium/drivers/virgl/ci/virgl-gl-skips.txt
+++ b/src/gallium/drivers/virgl/ci/virgl-gl-skips.txt
@ -5,7 +5,7 @@
 # Sometimes crashes, e.g. https://gitlab.freedesktop.org/kusma/mesa/-/jobs/4109419
 dEQP-GLES31.functional.compute.basic.empty

-# too slow.
+# too slow (>15s compile on host causes timeouts that make for flakes)
 dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36
 dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22

--- a/src/gallium/drivers/virgl/ci/virgl-gles-fails.txt
+++ b/src/gallium/drivers/virgl/ci/virgl-gles-fails.txt
@ -26,8 +26,11 @@ dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_rbo_2,Fail
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_texture_1,Fail
 dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_two_samples.multisample_texture_2,Fail
+
+# Times out waiting for >15s compile on the host side.
 dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36,Fail
 dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22,Fail
+
 KHR-GL30.glsl_noperspective.functionaltest,Fail

 KHR-GL30.transform_feedback.api_errors_test,Fail
--- a/src/gallium/drivers/virgl/ci/virgl-gles-flakes.txt
+++ b/src/gallium/drivers/virgl/ci/virgl-gles-flakes.txt
@ -12,9 +12,3 @@ dEQP-GLES2.functional.clipping.point.wide_point_clip_viewport_corner
 dEQP-GLES3.functional.clipping.point.wide_point_clip
 dEQP-GLES3.functional.clipping.point.wide_point_clip_viewport_center
 dEQP-GLES3.functional.clipping.point.wide_point_clip_viewport_corner
-
-dEQP-GLES31.functional.ssbo.layout.random.arrays_of_arrays.1
-
-dEQP-GLES31.functional.ssbo.layout.unsized_struct_array.single_buffer.std430_instance_array
-dEQP-GLES31.functional.ssbo.layout.unsized_struct_array.per_block_buffer.std430_instance_array
-dEQP-GLES31.functional.ssbo.layout.unsized_struct_array.single_buffer.std140_instance_array
--- a/src/gallium/drivers/virgl/ci/virgl-gles-skips.txt
+++ b/src/gallium/drivers/virgl/ci/virgl-gles-skips.txt
@ -4,5 +4,9 @@

 KHR-GL32.packed_pixels.varied_rectangle.depth*

+# too slow (>15s compile on host causes timeouts that make for flakes)
+dEQP-GLES31.functional.ssbo.layout.random.all_shared_buffer.36
+dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22
+
 # Sometimes crashes, e.g. https://gitlab.freedesktop.org/kusma/mesa/-/jobs/4109419
 dEQP-GLES31.functional.compute.basic.empty