add support for SPV_INTEL_subgroup_matrix_multiply_accumulate (#5928)

* add support for SPV_INTEL_subgroup_matrix_multiply_accumulate * Update DEPS --------- Co-authored-by: Alan Baker <alanbaker@google.com>
KhronosGroup · Jan 16, 2025 · f942f65 · f942f65
1 parent a6107ed
commit f942f65
Show file tree

Hide file tree

Showing 7 changed files with 45 additions and 4 deletions.
diff --git a/DEPS b/DEPS
@@ -14,7 +14,7 @@ vars = {
 
   're2_revision': '6dcd83d60f7944926bfd308cc13979fc53dd69ca',
 
-  'spirv_headers_revision': '0659679d9648a4dfdb5513efe25c495a3712dbf4',
+  'spirv_headers_revision': '2b2e05e088841c63c0b6fd4c9fb380d8688738d3',
 }
 
 deps = {

diff --git a/include/spirv-tools/libspirv.h b/include/spirv-tools/libspirv.h
@@ -320,6 +320,9 @@ typedef enum spv_operand_type_t {
   SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_REDUCE,
   // Enum type from SPV_NV_cooperative_matrix2
   SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS,
+  // Optional types from SPV_INTEL_subgroup_matrix_multiply_accumulate
+  SPV_OPERAND_TYPE_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS,
+  SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS,
 
   // This is a sentinel value, and does not represent an operand type.
   // It should come last.

diff --git a/source/binary.cpp b/source/binary.cpp
@@ -726,7 +726,9 @@ spv_result_t Parser::parseOperand(size_t inst_offset,
     case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_OPERANDS:
     case SPV_OPERAND_TYPE_OPTIONAL_COOPERATIVE_MATRIX_OPERANDS:
     case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_REDUCE:
-    case SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS: {
+    case SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS:
+    case SPV_OPERAND_TYPE_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS:
+    case SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS: {
       // This operand is a mask.
 
       // Map an optional operand type to its corresponding concrete type.
@@ -738,6 +740,9 @@ spv_result_t Parser::parseOperand(size_t inst_offset,
         parsed_operand.type = SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_OPERANDS;
       if (type == SPV_OPERAND_TYPE_OPTIONAL_RAW_ACCESS_CHAIN_OPERANDS)
         parsed_operand.type = SPV_OPERAND_TYPE_RAW_ACCESS_CHAIN_OPERANDS;
+      if (type == SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS)
+        parsed_operand.type =
+            SPV_OPERAND_TYPE_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS;
 
       // Check validity of set mask bits. Also prepare for operands for those
       // masks if they have any.  To get operand order correct, scan from

diff --git a/source/operand.cpp b/source/operand.cpp
@@ -237,6 +237,9 @@ const char* spvOperandTypeStr(spv_operand_type_t type) {
       return "cooperative matrix reduce";
     case SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS:
       return "tensor addressing operands";
+    case SPV_OPERAND_TYPE_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS:
+    case SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS:
+      return "matrix multiply accumulate operands";
     case SPV_OPERAND_TYPE_INITIALIZATION_MODE_QUALIFIER:
       return "initialization mode qualifier";
     case SPV_OPERAND_TYPE_HOST_ACCESS_QUALIFIER:
@@ -415,6 +418,7 @@ bool spvOperandIsConcreteMask(spv_operand_type_t type) {
     case SPV_OPERAND_TYPE_DEBUG_INFO_FLAGS:
     case SPV_OPERAND_TYPE_CLDEBUG100_DEBUG_INFO_FLAGS:
     case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_OPERANDS:
+    case SPV_OPERAND_TYPE_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS:
     case SPV_OPERAND_TYPE_RAW_ACCESS_CHAIN_OPERANDS:
     case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_REDUCE:
     case SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS:
@@ -437,6 +441,7 @@ bool spvOperandIsOptional(spv_operand_type_t type) {
     case SPV_OPERAND_TYPE_OPTIONAL_ACCESS_QUALIFIER:
     case SPV_OPERAND_TYPE_OPTIONAL_PACKED_VECTOR_FORMAT:
     case SPV_OPERAND_TYPE_OPTIONAL_COOPERATIVE_MATRIX_OPERANDS:
+    case SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS:
     case SPV_OPERAND_TYPE_OPTIONAL_CIV:
     case SPV_OPERAND_TYPE_OPTIONAL_RAW_ACCESS_CHAIN_OPERANDS:
     case SPV_OPERAND_TYPE_OPTIONAL_FPENCODING:

diff --git a/source/text.cpp b/source/text.cpp
@@ -416,7 +416,8 @@ spv_result_t spvTextEncodeOperand(const spvtools::AssemblyGrammar& grammar,
     case SPV_OPERAND_TYPE_CLDEBUG100_DEBUG_INFO_FLAGS:
     case SPV_OPERAND_TYPE_OPTIONAL_COOPERATIVE_MATRIX_OPERANDS:
     case SPV_OPERAND_TYPE_TENSOR_ADDRESSING_OPERANDS:
-    case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_REDUCE: {
+    case SPV_OPERAND_TYPE_COOPERATIVE_MATRIX_REDUCE:
+    case SPV_OPERAND_TYPE_OPTIONAL_MATRIX_MULTIPLY_ACCUMULATE_OPERANDS: {
       uint32_t value;
       if (auto error = grammar.parseMaskOperand(type, textValue, &value)) {
         return context->diagnostic(error)

diff --git a/test/binary_to_text_test.cpp b/test/binary_to_text_test.cpp
@@ -428,6 +428,33 @@ INSTANTIATE_TEST_SUITE_P(
                 "OpDecorate %1 HostAccessINTEL ReadWriteINTEL \"readwrite\"\n",
             })));
 
+// clang-format off
+INSTANTIATE_TEST_SUITE_P(
+    MatrixMultiplyAccumulateOperands, RoundTripInstructionsTest,
+    Combine(::testing::Values(SPV_ENV_UNIVERSAL_1_0),
+            ::testing::ValuesIn(std::vector<std::string>{
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 None\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixASignedComponentsINTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBSignedComponentsINTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixCBFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixResultBFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixAPackedInt8INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBPackedInt8INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixAPackedInt4INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBPackedInt4INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixATF32INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBTF32INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixCBFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixAPackedFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBPackedFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixAPackedBFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 MatrixBPackedBFloat16INTEL\n",
+                "%2 = OpSubgroupMatrixMultiplyAccumulateINTEL %1 %3 %4 %5 %6 "
+                "MatrixASignedComponentsINTEL|MatrixBSignedComponentsINTEL|MatrixAPackedInt8INTEL|MatrixBPackedInt8INTEL\n",
+            })));
+// clang-format on
+
 using MaskSorting = TextToBinaryTest;
 
 TEST_F(MaskSorting, MasksAreSortedFromLSBToMSB) {

diff --git a/utils/generate_grammar_tables.py b/utils/generate_grammar_tables.py
@@ -600,7 +600,7 @@ def generate_operand_kind_table(enums):
 
     # We have a few operand kinds that require their optional counterpart to
     # exist in the operand info table.
-    optional_enums = ['ImageOperands', 'AccessQualifier', 'MemoryAccess', 'PackedVectorFormat', 'CooperativeMatrixOperands', 'RawAccessChainOperands', 'FPEncoding']
+    optional_enums = ['ImageOperands', 'AccessQualifier', 'MemoryAccess', 'PackedVectorFormat', 'CooperativeMatrixOperands', 'MatrixMultiplyAccumulateOperands', 'RawAccessChainOperands', 'FPEncoding']
     optional_enums = [e for e in enums if e[0] in optional_enums]
     enums.extend(optional_enums)