Support transpose and pack for bit8 (#156065)

Valentine233 · pytorchmergebot · commit d26ca5de058d · 2025-07-07T01:40:47.000Z
To be used by CPU INT8 SDPA in torchao. pytorch/ao#2380 Pull Request resolved: #156065 Approved by: https://github.com/mingfeima, https://github.com/ezyang
diff --git a/aten/src/ATen/cpu/vec/vec_quant.h b/aten/src/ATen/cpu/vec/vec_quant.h
@@ -0,0 +1,153 @@
+#pragma once
+
+#include <ATen/cpu/vec/intrinsics.h>
+#include <c10/util/Exception.h>
+
+namespace at::vec {
+// See Note [CPU_CAPABILITY namespace]
+inline namespace CPU_CAPABILITY {
+
+// Transpose a [4, 64] block to [64, 4] (with contiguous output, ld=4)
+template <typename scalar_t, typename = std::enable_if_t<sizeof(scalar_t) == 1>>
+static inline void transpose_pad_4x64_block(
+    const scalar_t* src,
+    scalar_t* dst,
+    int64_t ld_src,
+    int krem = 4,
+    int nrem = 64) {
+#if defined(CPU_CAPABILITY_AVX512)
+  __m512i r[4];
+  // Load with mask if partial
+  if (nrem < 64) {
+    __mmask64 mask = (1ULL << nrem) - 1;
+    for (int i = 0; i < krem; ++i) {
+      r[i] = _mm512_maskz_loadu_epi8(mask, src + i * ld_src);
+    }
+    for (int i = krem; i < 4; ++i) {
+      r[i] = _mm512_setzero_si512();
+    }
+  } else {
+    for (int i = 0; i < krem; ++i) {
+      r[i] = _mm512_loadu_si512(
+          reinterpret_cast<const __m512i*>(src + i * ld_src));
+    }
+    for (int i = krem; i < 4; ++i) {
+      r[i] = _mm512_setzero_si512();
+    }
+  }
+
+  // Transpose 4x64 bytes using unpack and shuffle
+  __m512i t0 = _mm512_unpacklo_epi8(r[0], r[1]);
+  __m512i t1 = _mm512_unpackhi_epi8(r[0], r[1]);
+  __m512i t2 = _mm512_unpacklo_epi8(r[2], r[3]);
+  __m512i t3 = _mm512_unpackhi_epi8(r[2], r[3]);
+
+  __m512i u0 = _mm512_unpacklo_epi16(t0, t2);
+  __m512i u1 = _mm512_unpackhi_epi16(t0, t2);
+  __m512i u2 = _mm512_unpacklo_epi16(t1, t3);
+  __m512i u3 = _mm512_unpackhi_epi16(t1, t3);
+
+  __m512i v0 = _mm512_shuffle_i32x4(u0, u1, 0x88);
+  __m512i v1 = _mm512_shuffle_i32x4(u0, u1, 0xdd);
+  __m512i v2 = _mm512_shuffle_i32x4(u2, u3, 0x88);
+  __m512i v3 = _mm512_shuffle_i32x4(u2, u3, 0xdd);
+
+  __m512i r0 = _mm512_shuffle_i32x4(v0, v2, 0x88);
+  __m512i r1 = _mm512_shuffle_i32x4(v1, v3, 0x88);
+  __m512i r2 = _mm512_shuffle_i32x4(v0, v2, 0xdd);
+  __m512i r3 = _mm512_shuffle_i32x4(v1, v3, 0xdd);
+
+  // Store output
+  if (nrem < 16) {
+    __mmask64 mask = (1ULL << (nrem * 4)) - 1;
+    _mm512_mask_storeu_epi8(dst, mask, r0);
+  } else if (nrem == 16) {
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+  } else if (nrem < 32) {
+    int n_bytes1 = 64;
+    int n_bytes2 = (nrem * 4) - n_bytes1;
+    __mmask64 mask = (1ULL << n_bytes2) - 1;
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_mask_storeu_epi8(reinterpret_cast<__m512i*>(dst + 64), mask, r1);
+  } else if (nrem == 32) {
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64), r1);
+  } else if (nrem < 48) {
+    int n_bytes1 = 64 * 2;
+    int n_bytes2 = (nrem * 4) - n_bytes1;
+    __mmask64 mask = (1ULL << n_bytes2) - 1;
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64), r1);
+    _mm512_mask_storeu_epi8(reinterpret_cast<__m512i*>(dst + 64 * 2), mask, r2);
+  } else if (nrem == 48) {
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64), r1);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64 * 2), r2);
+  } else if (nrem < 64) {
+    int n_bytes1 = 64 * 3;
+    int n_bytes2 = (nrem * 4) - n_bytes1;
+    __mmask64 mask = (1ULL << n_bytes2) - 1;
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64), r1);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64 * 2), r2);
+    _mm512_mask_storeu_epi8(reinterpret_cast<__m512i*>(dst + 64 * 3), mask, r3);
+  } else {
+    // normal case, nrem == 64
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst), r0);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64), r1);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64 * 2), r2);
+    _mm512_storeu_si512(reinterpret_cast<__m512i*>(dst + 64 * 3), r3);
+  }
+#else
+  TORCH_CHECK(
+      false,
+      "transpose_pad_4x64_block is only supported when AVX-512 is supported")
+#endif
+}
+
+// Reorder [K, N] → [K/4, N, 4] (VNNI4-style layout for bit8)
+template <typename scalar_t, typename = std::enable_if_t<sizeof(scalar_t) == 1>>
+static inline void pack_vnni4(
+    const scalar_t* src,
+    scalar_t* dst,
+    int64_t ld_src,
+    int64_t K,
+    int64_t N) {
+#if defined(CPU_CAPABILITY_AVX512)
+  int64_t bk = 0;
+  int64_t _K = K / 4 * 4;
+  int64_t _N = N / 64 * 64;
+  for (; bk < _K; bk += 4) {
+    int64_t bn = 0;
+    for (; bn < _N; bn += 64) {
+      transpose_pad_4x64_block(
+          src + bk * ld_src + bn, dst + bk * N + bn * 4, ld_src);
+    }
+    int64_t nrem = N - bn;
+    if (nrem > 0) {
+      transpose_pad_4x64_block(
+          src + bk * ld_src + bn, dst + bk * N + bn * 4, ld_src, 4, nrem);
+    }
+  }
+
+  // Handle leftover K rows (< 4)
+  if (K % 4 != 0) {
+    int krem = K - bk;
+    int64_t bn = 0;
+    for (; bn < _N; bn += 64) {
+      transpose_pad_4x64_block(
+          src + bk * ld_src + bn, dst + bk * N + bn * 4, ld_src, krem);
+    }
+    int64_t nrem = N - bn;
+    if (nrem > 0) {
+      transpose_pad_4x64_block(
+          src + bk * ld_src + bn, dst + bk * N + bn * 4, ld_src, krem, nrem);
+    }
+  }
+#else
+  TORCH_CHECK(false, "pack_vnni4 is only supported when AVX-512 is supported")
+#endif
+}
+
+} // namespace CPU_CAPABILITY
+} // namespace at::vec
diff --git a/aten/src/ATen/native/cpu/utils.h b/aten/src/ATen/native/cpu/utils.h
@@ -165,6 +165,12 @@ inline void transpose<uint16_t>(int64_t M, int64_t N, const uint16_t* src, int64
   TORCH_CHECK(fbgemm::fbgemmSupportedCPU(), "Your CPU does not support FBGEMM.");
   fbgemm::transpose_simd<uint16_t>(M, N, src, ld_src, dst, ld_dst);
 }
+
+template <>
+inline void transpose<uint8_t>(int64_t M, int64_t N, const uint8_t* src, int64_t ld_src, uint8_t* dst, int64_t ld_dst) {
+  TORCH_CHECK(fbgemm::fbgemmSupportedCPU(), "Your CPU does not support FBGEMM.");
+  fbgemm::transpose_simd<uint8_t>(M, N, src, ld_src, dst, ld_dst);
+}
 #endif
 
 template <typename index_t, typename F>
diff --git a/aten/src/ATen/test/vec_test_all_types.cpp b/aten/src/ATen/test/vec_test_all_types.cpp
@@ -61,6 +61,8 @@ namespace {
     template <typename T>
     class QuantizationTests : public ::testing::Test {};
     template <typename T>
+    class Quantization8BitTests : public ::testing::Test {};
+    template <typename T>
     class Quantization8BitWithTailTests : public ::testing::Test {};
     template <typename T>
     class FunctionalTests : public ::testing::Test {};
@@ -79,6 +81,7 @@ namespace {
     using FloatTestedTypes = ::testing::Types<vfloat, vdouble, vcomplex, vcomplexDbl>;
     using ALLTestedTypes = ::testing::Types<vfloat, vdouble, vcomplex, vlong, vint, vshort, vqint8, vquint8, vqint>;
     using QuantTestedTypes = ::testing::Types<vqint8, vquint8, vqint>;
+    using Quantization8BitTestedTypes = ::testing::Types<vqint8, vquint8>;
 #if (defined(CPU_CAPABILITY_AVX2) ||  defined(CPU_CAPABILITY_AVX512))  && !defined(_MSC_VER)
     using Quantization8BitWithTailTestedTypes =
         ::testing::Types<vqint8, vquint8>;
@@ -116,6 +119,7 @@ namespace {
     TYPED_TEST_SUITE(BitwiseFloatsAdditional, RealFloatReducedFloatTestedTypes);
     TYPED_TEST_SUITE(BitwiseFloatsAdditional2, FloatTestedTypes);
     TYPED_TEST_SUITE(QuantizationTests, QuantTestedTypes);
+    TYPED_TEST_SUITE(Quantization8BitTests, Quantization8BitTestedTypes);
     TYPED_TEST_SUITE(InfiniteTests, RealFloatTestedTypes);
 #if (defined(CPU_CAPABILITY_AVX2) ||  defined(CPU_CAPABILITY_AVX512))  && !defined(_MSC_VER)
     TYPED_TEST_SUITE(
@@ -1496,6 +1500,68 @@ namespace {
             },
             test_case);
     }
+#ifndef _WIN32
+    TYPED_TEST(Quantization8BitTests, Transpose) {
+        using VT = ValueType<TypeParam>;
+        constexpr auto M = 4;
+        constexpr auto N = 64;
+        constexpr auto L = M * N;
+        constexpr auto ld_src = N;
+        constexpr auto ld_dst = M;
+        CACHE_ALIGN VT x[L];
+        CACHE_ALIGN VT y[L];
+        CACHE_ALIGN VT ref[L];
+        auto seed = TestSeed();
+        ValueGen<VT> generator(VT(-100), VT(100), seed);
+        for (const auto i : c10::irange(L)) {
+          x[i] = generator.get();
+        }
+        at::native::utils::transpose<uint8_t>(
+            M, N,
+            reinterpret_cast<uint8_t*>(x), ld_src,
+            reinterpret_cast<uint8_t*>(y), ld_dst);
+        for (int64_t j = 0; j < N; j++) {
+          for (int64_t i = 0; i < M; i++) {
+            ref[j * ld_dst + i] = c10::load(&(x[i * ld_src + j]));
+          }
+        }
+        for (const auto i : c10::irange(L)) {
+          ASSERT_EQ(y[i], ref[i])
+              << "Failure Details:\nTest Seed to reproduce: " << seed;
+        }
+    }
+#endif
+#if defined(CPU_CAPABILITY_AVX512)
+    TYPED_TEST(Quantization8BitTests, PackVNNI4) {
+        using VT = ValueType<TypeParam>;
+        constexpr auto K = 8;
+        constexpr auto N = 128;
+        constexpr auto L = K * N;
+        constexpr auto ld_src = N;
+        CACHE_ALIGN VT x[L];
+        CACHE_ALIGN VT y[L];
+        CACHE_ALIGN VT ref[L];
+        auto seed = TestSeed();
+        ValueGen<VT> generator(VT(-100), VT(100), seed);
+        for (const auto i : c10::irange(L)) {
+          x[i] = generator.get();
+        }
+        at::vec::pack_vnni4(x, y, ld_src, K, N);
+        int64_t _K = K / 4;
+        for (int64_t k = 0; k < _K; k++) {
+          for(int64_t n = 0; n < N; n++) {
+            for(int64_t l = 0; l < 4; l++) {
+              ref[k * N * 4 + n * 4 + l] =
+                  c10::load(&(x[k * ld_src * 4 + l * ld_src + n]));
+            }
+          }
+        }
+        for (const auto i : c10::irange(L)) {
+          ASSERT_EQ(y[i], ref[i])
+              << "Failure Details:\nTest Seed to reproduce: " << seed;
+        }
+    }
+#endif
     TYPED_TEST(FunctionalTests, Map) {
         using vec = TypeParam;
         using VT = ValueType<TypeParam>;
diff --git a/aten/src/ATen/test/vec_test_all_types.h b/aten/src/ATen/test/vec_test_all_types.h
@@ -1,6 +1,7 @@
 #pragma once
-#include <ATen/cpu/vec/vec.h>
 #include <ATen/cpu/vec/functional.h>
+#include <ATen/cpu/vec/vec.h>
+#include <ATen/cpu/vec/vec_quant.h>
 #include <c10/util/bit_cast.h>
 #include <c10/util/irange.h>
 #include <gtest/gtest.h>
@@ -21,7 +22,9 @@
 #else
 #define CACHE_LINE 32
 #endif
-
+#ifndef _WIN32
+#include <ATen/native/cpu/utils.h>
+#endif
 #if defined(__GNUC__)
 #define CACHE_ALIGN __attribute__((aligned(CACHE_LINE)))
 #define not_inline __attribute__((noinline))
