VERY experimental Nintendo Switch 2 emulator
C++
Avoid using VectorExtractDynamic for subgroup mask on Adreno GPUs
Billy Lawscfbe4b09eb30
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_warp.cppindex c5db19d098..77ff8c5731 100644--- a/src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp+++ b/src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp@@ -17,7 +17,22 @@ Id GetThreadId(EmitContext& ctx) {Id WarpExtract(EmitContext& ctx, Id value) {const Id thread_id{GetThreadId(ctx)};const Id local_index{ctx.OpShiftRightArithmetic(ctx.U32[1], thread_id, ctx.Const(5U))};- return ctx.OpVectorExtractDynamic(ctx.U32[1], value, local_index);+ if (ctx.profile.has_broken_spirv_subgroup_mask_vector_extract_dynamic) {+ const Id c0_sel{ctx.OpSelect(ctx.U32[1], ctx.OpIEqual(ctx.U1, local_index, ctx.Const(0U)),+ ctx.OpCompositeExtract(ctx.U32[1], value, 0U), ctx.Const(0U))};+ const Id c1_sel{ctx.OpSelect(ctx.U32[1], ctx.OpIEqual(ctx.U1, local_index, ctx.Const(1U)),+ ctx.OpCompositeExtract(ctx.U32[1], value, 1U), ctx.Const(0U))};+ const Id c2_sel{ctx.OpSelect(ctx.U32[1], ctx.OpIEqual(ctx.U1, local_index, ctx.Const(2U)),+ ctx.OpCompositeExtract(ctx.U32[1], value, 2U), ctx.Const(0U))};+ const Id c3_sel{ctx.OpSelect(ctx.U32[1], ctx.OpIEqual(ctx.U1, local_index, ctx.Const(3U)),+ ctx.OpCompositeExtract(ctx.U32[1], value, 3U), ctx.Const(0U))};+ const Id c0_or_c1{ctx.OpBitwiseOr(ctx.U32[1], c0_sel, c1_sel)};+ const Id c2_or_c3{ctx.OpBitwiseOr(ctx.U32[1], c2_sel, c3_sel)};+ const Id c0_or_c1_or_c2_or_c3{ctx.OpBitwiseOr(ctx.U32[1], c0_or_c1, c2_or_c3)};+ return c0_or_c1_or_c2_or_c3;+ } else {+ return ctx.OpVectorExtractDynamic(ctx.U32[1], value, local_index);+ }}Id LoadMask(EmitContext& ctx, Id mask) {diff --git a/src/shader_recompiler/profile.h b/src/shader_recompiler/profile.hindex 3bb4a7e6fc..9ca97f6a4b 100644--- a/src/shader_recompiler/profile.h+++ b/src/shader_recompiler/profile.h@@ -78,6 +78,8 @@ struct Profile {bool has_gl_bool_ref_bug{};/// Ignores SPIR-V ordered vs unordered using GLSL semanticsbool ignore_nan_fp_comparisons{};+ /// Some drivers have broken support for OpVectorExtractDynamic on subgroup mask inputs+ bool has_broken_spirv_subgroup_mask_vector_extract_dynamic{};u32 gl_max_compute_smem_size{};};diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cppindex e397137612..e506a8b301 100644--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp@@ -351,6 +351,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device.has_broken_signed_operations = false,.has_broken_fp16_float_controls = driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY,.ignore_nan_fp_comparisons = false,+ .has_broken_spirv_subgroup_mask_vector_extract_dynamic = driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY};host_info = Shader::HostTranslateInfo{.support_float16 = device.IsFloat16Supported(),