tint: Preserve padding in matrices with three rows

The PreservePadding transform now decomposes writes to matrices with
three rows into separate column vector writes, to avoid modifying
padding between columns.

Bug: tint:1571
Change-Id: If575f79bb87f52810783fd3338e2f3ce3228ab2e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/121600
Auto-Submit: James Price <jrprice@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: James Price <jrprice@google.com>
This commit is contained in:
James Price
2023-02-27 16:06:54 +00:00
committed by Dawn LUCI CQ
parent fe19fee3ea
commit 6176c85be8
84 changed files with 3733 additions and 1968 deletions

View File

@@ -12,12 +12,19 @@ layout(binding = 1, std430) buffer u_block_ssbo {
f16mat4x3 inner;
} s;
void assign_and_preserve_padding_s(f16mat4x3 value) {
s.inner[0] = value[0u];
s.inner[1] = value[1u];
s.inner[2] = value[2u];
s.inner[3] = value[3u];
}
f16mat4x3 load_u_inner() {
return f16mat4x3(u.inner_0, u.inner_1, u.inner_2, u.inner_3);
}
void f() {
s.inner = load_u_inner();
assign_and_preserve_padding_s(load_u_inner());
s.inner[1] = u.inner_0;
s.inner[1] = u.inner_0.zxy;
s.inner[0][1] = u.inner_1[0u];

View File

@@ -1,8 +1,15 @@
#include <metal_stdlib>
using namespace metal;
void assign_and_preserve_padding(device half4x3* const dest, half4x3 value) {
(*(dest))[0u] = value[0u];
(*(dest))[1u] = value[1u];
(*(dest))[2u] = value[2u];
(*(dest))[3u] = value[3u];
}
kernel void f(device half4x3* tint_symbol [[buffer(1)]], const constant half4x3* tint_symbol_1 [[buffer(0)]]) {
*(tint_symbol) = *(tint_symbol_1);
assign_and_preserve_padding(tint_symbol, *(tint_symbol_1));
(*(tint_symbol))[1] = (*(tint_symbol_1))[0];
(*(tint_symbol))[1] = half3((*(tint_symbol_1))[0]).zxy;
(*(tint_symbol))[0][1] = (*(tint_symbol_1))[1][0];

View File

@@ -1,7 +1,7 @@
; SPIR-V
; Version: 1.3
; Generator: Google Tint Compiler; 0
; Bound: 57
; Bound: 70
; Schema: 0
OpCapability Shader
OpCapability Float16
@@ -20,6 +20,8 @@
OpName %u_block "u_block"
OpMemberName %u_block 0 "inner"
OpName %s "s"
OpName %assign_and_preserve_padding_s "assign_and_preserve_padding_s"
OpName %value "value"
OpName %load_u_inner "load_u_inner"
OpName %f "f"
OpDecorate %u_block_std140 Block
@@ -45,53 +47,71 @@
%u_block = OpTypeStruct %mat4v3half
%_ptr_StorageBuffer_u_block = OpTypePointer StorageBuffer %u_block
%s = OpVariable %_ptr_StorageBuffer_u_block StorageBuffer
%10 = OpTypeFunction %mat4v3half
%void = OpTypeVoid
%10 = OpTypeFunction %void %mat4v3half
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%_ptr_Uniform_v3half = OpTypePointer Uniform %v3half
%uint_1 = OpConstant %uint 1
%uint_2 = OpConstant %uint 2
%uint_3 = OpConstant %uint 3
%void = OpTypeVoid
%33 = OpTypeFunction %void
%_ptr_StorageBuffer_mat4v3half = OpTypePointer StorageBuffer %mat4v3half
%int = OpTypeInt 32 1
%int_1 = OpConstant %int 1
%18 = OpConstantNull %int
%_ptr_StorageBuffer_v3half = OpTypePointer StorageBuffer %v3half
%50 = OpConstantNull %int
%21 = OpConstantNull %uint
%int_1 = OpConstant %int 1
%uint_1 = OpConstant %uint 1
%int_2 = OpConstant %int 2
%uint_2 = OpConstant %uint 2
%int_3 = OpConstant %int 3
%uint_3 = OpConstant %uint 3
%35 = OpTypeFunction %mat4v3half
%_ptr_Uniform_v3half = OpTypePointer Uniform %v3half
%53 = OpTypeFunction %void
%_ptr_StorageBuffer_half = OpTypePointer StorageBuffer %half
%53 = OpConstantNull %uint
%_ptr_Uniform_half = OpTypePointer Uniform %half
%load_u_inner = OpFunction %mat4v3half None %10
%12 = OpLabel
%18 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%19 = OpLoad %v3half %18
%22 = OpAccessChain %_ptr_Uniform_v3half %u %uint_1
%23 = OpLoad %v3half %22
%26 = OpAccessChain %_ptr_Uniform_v3half %u %uint_2
%27 = OpLoad %v3half %26
%30 = OpAccessChain %_ptr_Uniform_v3half %u %uint_3
%31 = OpLoad %v3half %30
%32 = OpCompositeConstruct %mat4v3half %19 %23 %27 %31
OpReturnValue %32
OpFunctionEnd
%f = OpFunction %void None %33
%36 = OpLabel
%38 = OpAccessChain %_ptr_StorageBuffer_mat4v3half %s %uint_0
%39 = OpFunctionCall %mat4v3half %load_u_inner
OpStore %38 %39
%43 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1
%44 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%45 = OpLoad %v3half %44
OpStore %43 %45
%46 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1
%47 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%48 = OpLoad %v3half %47
%49 = OpVectorShuffle %v3half %48 %48 2 0 1
OpStore %46 %49
%52 = OpAccessChain %_ptr_StorageBuffer_half %s %uint_0 %50 %int_1
%55 = OpAccessChain %_ptr_Uniform_half %u %uint_1 %53
%56 = OpLoad %half %55
OpStore %52 %56
%assign_and_preserve_padding_s = OpFunction %void None %10
%value = OpFunctionParameter %mat4v3half
%14 = OpLabel
%20 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %18
%22 = OpCompositeExtract %v3half %value 0
OpStore %20 %22
%24 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1
%26 = OpCompositeExtract %v3half %value 1
OpStore %24 %26
%28 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_2
%30 = OpCompositeExtract %v3half %value 2
OpStore %28 %30
%32 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_3
%34 = OpCompositeExtract %v3half %value 3
OpStore %32 %34
OpReturn
OpFunctionEnd
%load_u_inner = OpFunction %mat4v3half None %35
%37 = OpLabel
%41 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%42 = OpLoad %v3half %41
%44 = OpAccessChain %_ptr_Uniform_v3half %u %uint_1
%45 = OpLoad %v3half %44
%47 = OpAccessChain %_ptr_Uniform_v3half %u %uint_2
%48 = OpLoad %v3half %47
%50 = OpAccessChain %_ptr_Uniform_v3half %u %uint_3
%51 = OpLoad %v3half %50
%52 = OpCompositeConstruct %mat4v3half %42 %45 %48 %51
OpReturnValue %52
OpFunctionEnd
%f = OpFunction %void None %53
%55 = OpLabel
%57 = OpFunctionCall %mat4v3half %load_u_inner
%56 = OpFunctionCall %void %assign_and_preserve_padding_s %57
%58 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1
%59 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%60 = OpLoad %v3half %59
OpStore %58 %60
%61 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1
%62 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0
%63 = OpLoad %v3half %62
%64 = OpVectorShuffle %v3half %63 %63 2 0 1
OpStore %61 %64
%66 = OpAccessChain %_ptr_StorageBuffer_half %s %uint_0 %18 %int_1
%68 = OpAccessChain %_ptr_Uniform_half %u %uint_1 %21
%69 = OpLoad %half %68
OpStore %66 %69
OpReturn
OpFunctionEnd