[CPU] [Feat] Add native AMX-FP8 attention impl for Diamond Rapids (#49410)

Signed-off-by: zhejiangxiaomai <zhenhui.zhao@intel.com>
This commit is contained in:
zhao, zhenhui
2026-09-06 11:06:56 +08:00
committed by GitHub
parent 039ea82660
commit 1c344ed41e
10 changed files with 1211 additions and 61 deletions
+30 -2
View File
@@ -1,5 +1,17 @@
#include "cpu_attn_dispatch_generated.h"
#if defined(__x86_64__) || defined(_M_X64)
#include <cpuid.h>
// Runtime check for AMX-FP8 (Diamond Rapids): CPUID leaf 7, subleaf 1, EAX[21].
static bool runtime_has_amx_fp8() {
unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
if (__get_cpuid_count(7, 1, &eax, &ebx, &ecx, &edx)) {
return (eax >> 21) & 1u;
}
return false;
}
#endif // x86_64
// Maps kv_cache_dtype string to Fp8KVCacheDataType enum.
// "auto" -> kAuto(0); "fp8"/"fp8_e4m3" -> kFp8E4M3; "fp8_e5m2" -> kFp8E5M2.
static inline cpu_attention::Fp8KVCacheDataType parse_fp8_kv_dtype(
@@ -18,6 +30,16 @@ bool cpu_attn_has_isa(const std::string& isa) {
return true;
#else
return false;
#endif
}
if (isa == "amx_fp8") {
#ifdef CPU_CAPABILITY_AMXFP8
// Guard with runtime CPUID check: the binary may be compiled with
// -mamx-fp8 (CPU_CAPABILITY_AMXFP8 defined) but run on a CPU that
// has AMX-BF16 without AMX-FP8 (e.g. Sapphire Rapids, Emerald Rapids).
return runtime_has_amx_fp8();
#else
return false;
#endif
}
return false;
@@ -35,6 +57,8 @@ torch::Tensor get_scheduler_metadata(
cpu_attention::ISA isa;
if (isa_hint == "amx") {
isa = cpu_attention::ISA::AMX;
} else if (isa_hint == "amx_fp8") {
isa = cpu_attention::ISA::AMX_FP8;
} else if (isa_hint == "vec") {
isa = cpu_attention::ISA::VEC;
} else if (isa_hint == "vec16") {
@@ -130,6 +154,8 @@ void cpu_attn_reshape_and_cache(
cpu_attention::ISA isa_tag = [&]() {
if (isa == "amx") {
return cpu_attention::ISA::AMX;
} else if (isa == "amx_fp8") {
return cpu_attention::ISA::AMX_FP8;
} else if (isa == "vec") {
return cpu_attention::ISA::VEC;
} else if (isa == "vec16") {
@@ -149,8 +175,9 @@ void cpu_attn_reshape_and_cache(
if (is_fp8) {
TORCH_CHECK(isa_tag == cpu_attention::ISA::AMX ||
isa_tag == cpu_attention::ISA::AMX_FP8 ||
isa_tag == cpu_attention::ISA::VEC,
"FP8 KV cache is only supported on x86 (AMX/VEC) ISA");
"FP8 KV cache is only supported on x86 (AMX_FP8/AMX/VEC) ISA");
}
VLLM_DISPATCH_FLOATING_TYPES(
@@ -237,8 +264,9 @@ void cpu_attention_with_kv_cache(
input.k_scale_fp8 = static_cast<float>(k_scale);
input.v_scale_fp8 = static_cast<float>(v_scale);
TORCH_CHECK(input.metadata->isa == cpu_attention::ISA::AMX ||
input.metadata->isa == cpu_attention::ISA::AMX_FP8 ||
input.metadata->isa == cpu_attention::ISA::VEC,
"FP8 KV cache is only supported on x86 (AMX/VEC) ISA");
"FP8 KV cache is only supported on x86 (AMX_FP8/AMX/VEC) ISA");
}
VLLM_DISPATCH_FLOATING_TYPES(
+795
View File
@@ -0,0 +1,795 @@
// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
//
// AMX_FP8 attention implementation for Intel Diamond Rapids (DMR).
//
// Architecture:
// QK phase: native FP8×FP8 MMA via _tile_dpfp8ps / _tile_dpbf8ps
// Q is quantized from BF16 to the KV cache FP8 type.
// K cache is stored in the native AMX-FP8 layout (1 byte/element).
// PV phase: native FP8 MMA with 64-byte VNNI4 tiles.
// Softmax writes P once in compact FP8 form for reuse across all
// output-dimension groups.
//
// Scale bookkeeping for QK:
// true_score = Q_fp32 · K_fp32
// = (Q_fp8 * q_scale) · (K_fp8 * k_scale)
// = dot(Q_fp8, K_fp8) * q_scale * k_scale
// execute_attention() applies: scale *= q_dynamic_scale * k_scale
// where q_dynamic_scale is computed per-tile in copy_q_heads_tile().
//
// Scale bookkeeping for PV:
// raw_C = dot(P_fp8, V_fp8)
// ≈ dot(P/p_scale, V_real/v_scale)
// = dot(P, V_real) / (p_scale * v_scale)
// In final_output: multiply by output_v_scale = p_scale * v_scale.
//
#ifndef CPU_ATTN_AMX_FP8_HPP
#define CPU_ATTN_AMX_FP8_HPP
#ifdef CPU_CAPABILITY_AMXFP8
#include "cpu_attn_amx.hpp" // pulls in TileGemm224/TileGemm122 BF16 specialisations
#include "cpu_attn_fp8.hpp"
#include "cpu_attn_impl.hpp"
#include "cpu_types.hpp"
#include <amxfp8intrin.h> // requires GCC 14+ / Clang 17+, -mamx-fp8
namespace cpu_attention {
// ---------------------------------------------------------------------------
// Native-FP8 K cache layout reshape helper.
//
// Layout per 16-token group (token_num_per_group = AMX_TILE_ROW_NUM = 16):
// k-slice s (s = 0 .. head_dim/64 - 1):
// row k (k = 0..15): four consecutive K values for each of 16 tokens
// (VNNI4 B-operand layout)
// One k-slice = 16 rows × 64 bytes = 1024 bytes = 1 AMX FP8 tile.
// Full group = (head_dim / 64) slices × 1024 bytes.
//
// For head_dim=128: group = 2 × 1024 = 2048 bytes.
// Consecutive 16-token groups are contiguous within each cache block.
// ---------------------------------------------------------------------------
template <typename scalar_t, uint8_t (*quant_fn)(float, float)>
inline void reshape_and_cache_k_amx_fp8_impl(
const scalar_t* key_ptr, uint8_t* key_cache_ptr, const int64_t* slot_ptr,
int64_t token_num, int64_t head_num, int64_t head_dim, int64_t block_size,
int64_t k_stride0, int64_t k_stride1, int64_t kc_stride0,
int64_t kc_stride1, float k_inv) {
constexpr int64_t token_num_per_group = AMX_TILE_ROW_NUM; // 16
constexpr int64_t fp8_per_row = AMX_TILE_ROW_BYTES; // 64 FP8/row
const int64_t slice_num = head_dim / fp8_per_row;
// Bytes per 16-token group = slice_num * (16 * 64) = head_dim * 16
const int64_t group_bytes = token_num_per_group * head_dim;
#pragma omp parallel for collapse(2) schedule(static)
for (int64_t tok = 0; tok < token_num; ++tok) {
for (int64_t h = 0; h < head_num; ++h) {
const int64_t slot = slot_ptr[tok];
if (slot < 0) continue;
const int64_t block_idx = slot / block_size;
const int64_t block_offset = slot % block_size;
const int64_t group_idx = block_offset / token_num_per_group;
const int64_t group_offset = block_offset % token_num_per_group;
const scalar_t* ksrc = key_ptr + tok * k_stride0 + h * k_stride1;
uint8_t* group_base = key_cache_ptr + block_idx * kc_stride0 +
h * kc_stride1 + group_idx * group_bytes;
// Each k-slice is one 16x64-byte VNNI4 B tile. Row k stores dimensions
// [4*k, 4*k+4) for every token; each token occupies four bytes.
for (int64_t s = 0; s < slice_num; ++s) {
uint8_t* tile_base = group_base + s * token_num_per_group * fp8_per_row;
for (int64_t k = 0; k < fp8_per_row / 4; ++k) {
uint8_t* packed_values =
tile_base + k * fp8_per_row + group_offset * 4;
for (int64_t lane = 0; lane < 4; ++lane) {
packed_values[lane] = quant_fn(
static_cast<float>(ksrc[s * fp8_per_row + k * 4 + lane]),
k_inv);
}
}
}
}
}
}
template <typename fp8_t>
__attribute__((target("avx10.2"), noinline)) void quantize_probability_row(
const float* __restrict__ src, uint8_t* __restrict__ dst, int32_t cols) {
constexpr float fp8_max =
std::is_same_v<fp8_t, c10::Float8_e4m3fn> ? 448.0f : 57344.0f;
const __m512 scale = _mm512_set1_ps(fp8_max);
for (int32_t col = 0; col < cols; col += 16) {
const __m512 scaled = _mm512_mul_ps(_mm512_loadu_ps(src + col), scale);
const __m256h fp16 = (__m256h)_mm512_cvtps_ph(
scaled, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
if constexpr (std::is_same_v<fp8_t, c10::Float8_e4m3fn>) {
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + col),
_mm256_cvtph_hf8(fp16));
} else {
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + col),
_mm256_cvtph_bf8(fp16));
}
}
}
template <typename fp8_t>
__attribute__((target("avx10.2"), noinline)) void quantize_q_tile(
const c10::BFloat16* __restrict__ src, uint8_t* __restrict__ dst,
int32_t q_num, int32_t q_heads_per_kv, int64_t q_num_stride,
int64_t q_head_stride, int64_t head_dim, float inv_scale) {
const __m512 scale = _mm512_set1_ps(inv_scale);
constexpr float fp8_max =
std::is_same_v<fp8_t, c10::Float8_e4m3fn> ? 448.0f : 57344.0f;
const __m512 max_value = _mm512_set1_ps(fp8_max);
const __m512 min_value = _mm512_set1_ps(-fp8_max);
const int64_t head_size_block_num = head_dim / AMX_TILE_ROW_BYTES;
int32_t packed_row = 0;
for (int32_t q = 0; q < q_num; ++q) {
const c10::BFloat16* head = src + q * q_num_stride;
for (int32_t h = 0; h < q_heads_per_kv;
++h, head += q_head_stride, ++packed_row) {
uint8_t* packed_head =
dst +
(packed_row / AMX_TILE_ROW_NUM) * head_size_block_num *
AMX_TILE_BYTES +
(packed_row % AMX_TILE_ROW_NUM) * AMX_TILE_ROW_BYTES;
for (int64_t e = 0; e < head_dim; e += 16) {
const __m256i bf16 =
_mm256_loadu_si256(reinterpret_cast<const __m256i*>(head + e));
__m512 values = _mm512_castsi512_ps(
_mm512_slli_epi32(_mm512_cvtepu16_epi32(bf16), 16));
values = _mm512_mul_ps(values, scale);
values = _mm512_min_ps(values, max_value);
values = _mm512_max_ps(values, min_value);
const __m256h fp16 = (__m256h)_mm512_cvtps_ph(
values, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
const int64_t block = e / AMX_TILE_ROW_BYTES;
const int64_t offset = e % AMX_TILE_ROW_BYTES;
__m128i fp8;
if constexpr (std::is_same_v<fp8_t, c10::Float8_e4m3fn>) {
fp8 = _mm256_cvtph_hf8(fp16);
} else {
fp8 = _mm256_cvtph_bf8(fp16);
}
_mm_storeu_si128(reinterpret_cast<__m128i*>(
packed_head + block * AMX_TILE_BYTES + offset),
fp8);
}
}
}
}
// ---------------------------------------------------------------------------
// ---------------------------------------------------------------------------
// TileGemm224<uint8_t, kv_cache_t>: 2-2-4 pattern for AMX_FP8.
//
// QK phase: native FP8 × FP8 MMA (_tile_dphf8ps / _tile_dpbf8ps).
// PV phase: native wide FP8 MMA (colsb=64 tiles).
// P is quantized once per KV tile and reused by every output-dimension group.
// V (FP8 E4M3) loaded directly from cache (no dequant).
// Result correction is folded into the final output scale.
// ---------------------------------------------------------------------------
template <typename kv_cache_t>
class TileGemm224<uint8_t, kv_cache_t> {
static_assert(std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
std::is_same_v<kv_cache_t, c10::Float8_e5m2>,
"kv_cache_t must be Float8_e4m3fn or Float8_e5m2");
// FP8 tile size in uint8_t elements (= AMX_TILE_BYTES, one per byte).
static constexpr int64_t fp8_tile_elems = AMX_TILE_BYTES; // 1024
public:
static constexpr bool prequantize_probabilities = true;
FORCE_INLINE static void quantize_probability_row(const float* src,
uint8_t* dst,
int32_t cols) {
cpu_attention::quantize_probability_row<kv_cache_t>(src, dst, cols);
}
// ------------------------------------------------------------------
// QK: k_times = head_dim / 64 (each FP8 tile covers 64 k-elements).
// PV: native wide FP8 MMA.
// ------------------------------------------------------------------
template <AttentionGemmPhase phase, int32_t k_size>
FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
void* __restrict__ b_tile,
float* __restrict__ c_tile, const int64_t lda,
const int64_t ldb, const int64_t ldc,
const int32_t block_size,
const int32_t dynamic_k_size,
const bool accum_c) {
if constexpr (phase == AttentionGemmPhase::PV) {
// -------- Native wide FP8 PV (FP8 × FP8 → FP32) --------
//
// A fixed FP8-dependent P scale lets the tile accumulate all k-steps
// before get_output_v_scale() applies the correction.
const int32_t m_0 =
std::min(m_size, static_cast<int32_t>(AMX_TILE_ROW_NUM));
const int32_t m_1 = m_size - m_0;
constexpr int32_t v_stride = AMX_TILE_ROW_BYTES;
const int32_t k_times =
dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(uint8_t));
constexpr int32_t a_advance = AMX_TILE_ROW_BYTES;
constexpr int32_t v_advance = AMX_TILE_BYTES;
const int32_t c_stride = ldc * sizeof(float);
for (int32_t output_group = 0; output_group < 2; ++output_group) {
uint8_t* a0 = reinterpret_cast<uint8_t*>(a_tile);
uint8_t* a1 = a0 + lda * AMX_TILE_ROW_NUM;
uint8_t* v2 = reinterpret_cast<uint8_t*>(b_tile) +
output_group * 2 * block_size * AMX_TILE_ROW_BYTES / 4;
uint8_t* v3 = v2 + block_size * AMX_TILE_ROW_BYTES / 4;
float* c4 = c_tile + output_group * 2 * AMX_TILE_ROW_NUM;
float* c5 = c4 + AMX_TILE_ROW_NUM;
float* c6 = c4 + AMX_TILE_ROW_NUM * ldc;
float* c7 = c6 + AMX_TILE_ROW_NUM;
if (accum_c) {
_tile_loadd(4, c4, c_stride);
_tile_loadd(5, c5, c_stride);
if (m_1 > 0) {
_tile_loadd(6, c6, c_stride);
_tile_loadd(7, c7, c_stride);
}
} else {
_tile_zero(4);
_tile_zero(5);
if (m_1 > 0) {
_tile_zero(6);
_tile_zero(7);
}
}
for (int32_t k = 0; k < k_times; ++k) {
_tile_loadd(0, a0, lda);
_tile_stream_loadd(2, v2, v_stride);
_tile_stream_loadd(3, v3, v_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(4, 0, 2);
_tile_dphf8ps(5, 0, 3);
} else {
_tile_dpbf8ps(4, 0, 2);
_tile_dpbf8ps(5, 0, 3);
}
if (m_1 > 0) {
_tile_loadd(1, a1, lda);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 1, 2);
_tile_dphf8ps(7, 1, 3);
} else {
_tile_dpbf8ps(6, 1, 2);
_tile_dpbf8ps(7, 1, 3);
}
}
a0 += a_advance;
a1 += a_advance;
v2 += v_advance;
v3 += v_advance;
}
_tile_stored(4, c4, c_stride);
_tile_stored(5, c5, c_stride);
if (m_1 > 0) {
_tile_stored(6, c6, c_stride);
_tile_stored(7, c7, c_stride);
}
}
return;
}
// -------- QK phase: native FP8 MMA --------
// k_times: each FP8 AMX tile covers (AMX_TILE_ROW_NUM * 4) FP8 = 64
// inner-product elements.
const int32_t k_times =
dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(uint8_t));
const int32_t b_tile_stride = AMX_TILE_ROW_BYTES; // 64 bytes
const int32_t m_0 =
std::min(m_size, static_cast<int32_t>(AMX_TILE_ROW_NUM));
const int32_t m_1 = m_size - m_0;
const int32_t c_tile_stride = ldc * sizeof(float);
constexpr int32_t output_group_num = 2;
for (int32_t output_group = 0; output_group < output_group_num;
++output_group) {
uint8_t* __restrict__ a_tile_0 = static_cast<uint8_t*>(a_tile);
uint8_t* __restrict__ a_tile_1 = a_tile_0 + k_times * fp8_tile_elems;
uint8_t* __restrict__ b_tile_2 =
static_cast<uint8_t*>(b_tile) +
output_group * 2 * k_size * AMX_TILE_ROW_NUM;
uint8_t* __restrict__ b_tile_3 = b_tile_2 + (k_size * AMX_TILE_ROW_NUM);
float* __restrict__ c_tile_4 = c_tile + output_group * 32;
float* __restrict__ c_tile_5 = c_tile_4 + AMX_TILE_ROW_NUM;
float* __restrict__ c_tile_6 = c_tile_4 + AMX_TILE_ROW_NUM * ldc;
float* __restrict__ c_tile_7 = c_tile_6 + AMX_TILE_ROW_NUM;
if (accum_c) {
_tile_loadd(4, c_tile_4, c_tile_stride);
_tile_loadd(5, c_tile_5, c_tile_stride);
_tile_loadd(6, c_tile_6, c_tile_stride);
_tile_loadd(7, c_tile_7, c_tile_stride);
} else {
_tile_zero(4);
_tile_zero(5);
_tile_zero(6);
_tile_zero(7);
}
for (int32_t k = 0; k < k_times; ++k) {
_tile_loadd(0, a_tile_0, AMX_TILE_ROW_BYTES);
_tile_stream_loadd(2, b_tile_2, b_tile_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(4, 0, 2);
} else {
_tile_dpbf8ps(4, 0, 2);
}
_tile_stream_loadd(3, b_tile_3, b_tile_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(5, 0, 3);
} else {
_tile_dpbf8ps(5, 0, 3);
}
_tile_loadd(1, a_tile_1, AMX_TILE_ROW_BYTES);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 1, 2);
_tile_dphf8ps(7, 1, 3);
} else {
_tile_dpbf8ps(6, 1, 2);
_tile_dpbf8ps(7, 1, 3);
}
// Q is packed as [row group][K slice].
a_tile_0 += fp8_tile_elems;
a_tile_1 += fp8_tile_elems;
// Advance K cache (native layout): 1 FP8 tile per k-step.
b_tile_2 += fp8_tile_elems;
b_tile_3 += fp8_tile_elems;
}
_tile_stored(4, c_tile_4, c_tile_stride);
_tile_stored(5, c_tile_5, c_tile_stride);
_tile_stored(6, c_tile_6, c_tile_stride);
_tile_stored(7, c_tile_7, c_tile_stride);
}
}
FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
// QK and PV share one 64-byte tile configuration.
const int32_t m_0 = std::min(m, static_cast<int32_t>(AMX_TILE_ROW_NUM));
const int32_t m_1 = m - m_0;
// Use all 64 bytes per row for every tile (same as BF16 AMX).
// This works for both FP8 (64 FP8/row) and BF16 (32 BF16/row = 64B).
for (int i = 0; i < 8; ++i) config.colsb[i] = AMX_TILE_ROW_BYTES;
config.rows[0] = m_0;
config.rows[1] = m_1;
config.rows[2] = AMX_TILE_ROW_NUM;
config.rows[3] = AMX_TILE_ROW_NUM;
config.rows[4] = m_0;
config.rows[5] = m_0;
config.rows[6] = m_1;
config.rows[7] = m_1;
_tile_loadconfig(&config);
}
};
// ---------------------------------------------------------------------------
// TileGemm122<uint8_t, kv_cache_t>: 1-2-2 pattern for AMX_FP8.
// ---------------------------------------------------------------------------
template <typename kv_cache_t>
class TileGemm122<uint8_t, kv_cache_t> {
static_assert(std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
std::is_same_v<kv_cache_t, c10::Float8_e5m2>,
"kv_cache_t must be Float8_e4m3fn or Float8_e5m2");
static constexpr int64_t fp8_tile_elems = AMX_TILE_BYTES;
public:
static constexpr bool prequantize_probabilities = true;
FORCE_INLINE static void quantize_probability_row(const float* src,
uint8_t* dst,
int32_t cols) {
cpu_attention::quantize_probability_row<kv_cache_t>(src, dst, cols);
}
template <AttentionGemmPhase phase, int32_t k_size>
FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
void* __restrict__ b_tile,
float* __restrict__ c_tile, const int64_t lda,
const int64_t ldb, const int64_t ldc,
const int32_t block_size,
const int32_t dynamic_k_size,
const bool accum_c) {
if constexpr (phase == AttentionGemmPhase::PV) {
// Native wide FP8 PV (1-2-2 pattern)
// A fixed FP8-dependent P scale enables accumulation across all k-steps.
constexpr int32_t v_stride = AMX_TILE_ROW_BYTES;
constexpr int32_t a_advance = AMX_TILE_ROW_BYTES;
constexpr int32_t v_advance = AMX_TILE_BYTES;
const int32_t k_times =
dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(uint8_t));
uint8_t* a = reinterpret_cast<uint8_t*>(a_tile);
uint8_t* v2 = reinterpret_cast<uint8_t*>(b_tile);
uint8_t* v3 = v2 + (block_size * AMX_TILE_ROW_BYTES / 4);
uint8_t* v4 = v3 + (block_size * AMX_TILE_ROW_BYTES / 4);
uint8_t* v5 = v4 + (block_size * AMX_TILE_ROW_BYTES / 4);
float* c4 = c_tile + 2 * AMX_TILE_ROW_NUM;
float* c5 = c_tile + 3 * AMX_TILE_ROW_NUM;
float* c6 = c_tile;
float* c7 = c6 + AMX_TILE_ROW_BYTES / sizeof(float);
const int32_t c_stride = ldc * sizeof(float);
if (accum_c) {
_tile_loadd(4, c4, c_stride);
_tile_loadd(5, c5, c_stride);
_tile_loadd(6, c6, c_stride);
_tile_loadd(7, c7, c_stride);
} else {
_tile_zero(4);
_tile_zero(5);
_tile_zero(6);
_tile_zero(7);
}
for (int32_t k = 0; k < k_times; ++k) {
_tile_loadd(0, a, lda);
_tile_stream_loadd(2, v2, v_stride);
_tile_stream_loadd(3, v3, v_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 0, 2);
_tile_dphf8ps(7, 0, 3);
} else {
_tile_dpbf8ps(6, 0, 2);
_tile_dpbf8ps(7, 0, 3);
}
_tile_stream_loadd(2, v4, v_stride);
_tile_stream_loadd(3, v5, v_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(4, 0, 2);
_tile_dphf8ps(5, 0, 3);
} else {
_tile_dpbf8ps(4, 0, 2);
_tile_dpbf8ps(5, 0, 3);
}
a += a_advance;
v2 += v_advance;
v3 += v_advance;
v4 += v_advance;
v5 += v_advance;
}
_tile_stored(4, c4, c_stride);
_tile_stored(5, c5, c_stride);
_tile_stored(6, c6, c_stride);
_tile_stored(7, c7, c_stride);
return;
}
// -------- QK phase: native FP8 MMA (1-2-2 pattern) --------
const int32_t k_times =
dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(uint8_t));
const int32_t k_group_times = k_times / 2;
const bool has_tail = (k_times % 2 == 1);
const int32_t b_stride = AMX_TILE_ROW_BYTES;
const int32_t c_stride = ldc * sizeof(float);
constexpr int32_t output_group_num = 2;
for (int32_t output_group = 0; output_group < output_group_num;
++output_group) {
uint8_t* __restrict__ a_tile_0 = static_cast<uint8_t*>(a_tile);
uint8_t* __restrict__ a_tile_1 = a_tile_0 + fp8_tile_elems;
uint8_t* __restrict__ b_tile_2 =
static_cast<uint8_t*>(b_tile) +
output_group * 2 * k_size * AMX_TILE_ROW_NUM;
uint8_t* __restrict__ b_tile_3 = b_tile_2 + (k_size * AMX_TILE_ROW_NUM);
uint8_t* __restrict__ b_tile_4 = b_tile_2 + fp8_tile_elems;
uint8_t* __restrict__ b_tile_5 = b_tile_3 + fp8_tile_elems;
float* __restrict__ c_tile_6 = c_tile + output_group * 32;
float* __restrict__ c_tile_7 = c_tile_6 + AMX_TILE_ROW_NUM;
if (accum_c) {
_tile_loadd(6, c_tile_6, c_stride);
_tile_loadd(7, c_tile_7, c_stride);
} else {
_tile_zero(6);
_tile_zero(7);
}
for (int32_t k = 0; k < k_group_times; ++k) {
_tile_loadd(0, a_tile_0, AMX_TILE_ROW_BYTES);
_tile_stream_loadd(2, b_tile_2, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 0, 2);
} else {
_tile_dpbf8ps(6, 0, 2);
}
_tile_stream_loadd(3, b_tile_3, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(7, 0, 3);
} else {
_tile_dpbf8ps(7, 0, 3);
}
_tile_loadd(1, a_tile_1, AMX_TILE_ROW_BYTES);
_tile_stream_loadd(2, b_tile_4, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 1, 2);
} else {
_tile_dpbf8ps(6, 1, 2);
}
_tile_stream_loadd(3, b_tile_5, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(7, 1, 3);
} else {
_tile_dpbf8ps(7, 1, 3);
}
a_tile_0 += 2 * fp8_tile_elems;
a_tile_1 += 2 * fp8_tile_elems;
b_tile_2 += 2 * fp8_tile_elems;
b_tile_3 += 2 * fp8_tile_elems;
b_tile_4 += 2 * fp8_tile_elems;
b_tile_5 += 2 * fp8_tile_elems;
}
if (has_tail) {
_tile_loadd(0, a_tile_0, AMX_TILE_ROW_BYTES);
_tile_stream_loadd(2, b_tile_2, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(6, 0, 2);
} else {
_tile_dpbf8ps(6, 0, 2);
}
_tile_stream_loadd(3, b_tile_3, b_stride);
if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
_tile_dphf8ps(7, 0, 3);
} else {
_tile_dpbf8ps(7, 0, 3);
}
}
_tile_stored(6, c_tile_6, c_stride);
_tile_stored(7, c_tile_7, c_stride);
}
}
FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
for (int i = 0; i < 8; ++i) config.colsb[i] = AMX_TILE_ROW_BYTES;
config.rows[0] = m;
config.rows[1] = m;
config.rows[2] = AMX_TILE_ROW_NUM;
config.rows[3] = AMX_TILE_ROW_NUM;
config.rows[4] = m;
config.rows[5] = m;
config.rows[6] = m;
config.rows[7] = m;
_tile_loadconfig(&config);
}
};
// ---------------------------------------------------------------------------
// AttentionImpl<ISA::AMX_FP8, scalar_t, head_dim, kv_cache_scalar_t>
//
// scalar_t = BF16 (input query/output type)
// kv_cache_scalar_t = Float8_e4m3fn or Float8_e5m2
// ---------------------------------------------------------------------------
template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
requires(std::is_same_v<scalar_t, c10::BFloat16> &&
(std::is_same_v<kv_cache_scalar_t, c10::Float8_e4m3fn> ||
std::is_same_v<kv_cache_scalar_t, c10::Float8_e5m2>) &&
(head_dim % 64 == 0))
class AttentionImpl<ISA::AMX_FP8, scalar_t, head_dim, kv_cache_scalar_t> {
public:
// Q buffer holds FP8-quantized query.
using query_t = scalar_t; // BF16 input
using q_buffer_t = uint8_t; // FP8 quantized, stored in Q buffer
using kv_cache_t = kv_cache_scalar_t;
using logits_buffer_t = float;
using partial_output_buffer_t = float;
using prob_buffer_t = scalar_t; // BF16 softmax probs (PV phase)
constexpr static int64_t BlockSizeAlignment = 64;
constexpr static int64_t HeadDimAlignment = 64;
constexpr static int64_t MaxQHeadNumPerIteration = 32;
constexpr static int64_t HeadDim = head_dim;
constexpr static ISA ISAType = ISA::AMX_FP8;
constexpr static bool scale_on_logits = true;
float k_scale = 1.0f;
float v_scale = 1.0f;
// Dynamically computed per copy_q_heads_tile() call.
float q_dynamic_scale = 1.0f;
public:
AttentionImpl() : current_q_head_num_(0) {
vec_op::unroll_loop<int, 8>([&](int i) { amx_tile_config_.colsb[i] = 64; });
}
~AttentionImpl() { _tile_release(); }
void init_from_input(const AttentionInput* input) {
k_scale = input->k_scale_fp8;
v_scale = input->v_scale_fp8;
// q_scale_fp8 from input is used as a hint; dynamic scale is recomputed
// per-tile in copy_q_heads_tile().
}
// Output scale for the PV result.
float get_output_v_scale() const noexcept {
constexpr float fp8_max =
std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ? 448.0f : 57344.0f;
return v_scale / fp8_max;
}
template <template <typename tile_gemm_t> typename attention>
FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
// Apply QK scale correction: score_fp32 = dot(Q_fp8, K_fp8) * q_dyn * k
scale *= q_dynamic_scale * k_scale;
if (q_head_num > AMX_TILE_ROW_NUM) {
if (q_head_num != current_q_head_num_) {
current_q_head_num_ = q_head_num;
TileGemm224<q_buffer_t, kv_cache_t>::init_tile_config(q_head_num,
amx_tile_config_);
}
attention<TileGemm224<q_buffer_t, kv_cache_t>> attention_iteration;
attention_iteration(CPU_ATTENTION_PARAMS);
} else {
if (q_head_num != current_q_head_num_) {
current_q_head_num_ = q_head_num;
TileGemm122<q_buffer_t, kv_cache_t>::init_tile_config(q_head_num,
amx_tile_config_);
}
attention<TileGemm122<q_buffer_t, kv_cache_t>> attention_iteration;
attention_iteration(CPU_ATTENTION_PARAMS);
}
}
// K cache stride: same element-count formula as AMX_BF16.
// (BlockSizeAlignment * head_dim) uint8_t elements = same byte count as
// before since kv_cache_t is 1-byte FP8.
constexpr static int64_t k_cache_token_group_stride(
const int32_t /*block_size*/) {
return BlockSizeAlignment * head_dim;
}
// Both native FP8 formats use VNNI4 packing.
constexpr static int64_t v_cache_token_group_stride(
const int32_t /*block_size*/) {
return BlockSizeAlignment * (AMX_TILE_ROW_BYTES / 4);
}
constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
return block_size * HeadDimAlignment;
}
// Copy Q heads tile: quantize BF16 to the KV cache FP8 type and pack it.
//
// Q buffer layout (FP8):
// Same tile-prepacking as AMX_BF16 copy_q_heads_tile but with uint8_t
// (FP8) elements. Each FP8 tile row = 64 bytes = 64 FP8 elements.
// head_size_block_num = head_dim / 64 (vs head_dim/32 for BF16).
//
// Dynamic scale: max(abs(Q)) / fp8_max is computed over all elements,
// stored in q_dynamic_scale for use in execute_attention().
void copy_q_heads_tile(
scalar_t* __restrict__ src, // [q_num, q_heads_per_kv, head_size] BF16
q_buffer_t* __restrict__ q_buffer, const int32_t q_num,
const int32_t q_heads_per_kv, const int64_t q_num_stride,
const int64_t q_head_stride, const float /*scale*/) {
// ---- Step 1: compute dynamic per-tile Q scale ----
constexpr float fp8_max =
std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ? 448.0f : 57344.0f;
float max_abs = 0.0f;
{
scalar_t* src_iter = src;
__m512 max_abs_vec = _mm512_setzero_ps();
const __m512i abs_mask = _mm512_set1_epi32(0x7FFFFFFF);
for (int32_t q = 0; q < q_num; ++q, src_iter += q_num_stride) {
scalar_t* head_iter = src_iter;
for (int32_t h = 0; h < q_heads_per_kv;
++h, head_iter += q_head_stride) {
for (int64_t e = 0; e < head_dim; e += 16) {
const __m256i bf16 = _mm256_loadu_si256(
reinterpret_cast<const __m256i*>(head_iter + e));
const __m512i fp32_bits =
_mm512_slli_epi32(_mm512_cvtepu16_epi32(bf16), 16);
const __m512 abs_values =
_mm512_castsi512_ps(_mm512_and_si512(fp32_bits, abs_mask));
max_abs_vec = _mm512_max_ps(max_abs_vec, abs_values);
}
}
}
max_abs = _mm512_reduce_max_ps(max_abs_vec);
}
if (max_abs == 0.0f) max_abs = 1.0f; // avoid div-by-zero
const float inv_scale = fp8_max / max_abs;
// Store true scale for score correction in execute_attention().
q_dynamic_scale = max_abs / fp8_max;
// ---- Step 2: quantize BF16 Q to FP8 and pack into q_buffer ----
// Layout: same tile-prepacking as AMX_BF16, but with FP8 (1 byte/elem).
// head_size_block_num = head_dim * sizeof(uint8_t) / AMX_TILE_ROW_BYTES
// = head_dim / 64
constexpr int64_t bytes_per_head = head_dim * sizeof(uint8_t);
static_assert(bytes_per_head % AMX_TILE_ROW_BYTES == 0,
"head_dim must be divisible by 64 for AMX_FP8");
quantize_q_tile<kv_cache_t>(src, reinterpret_cast<uint8_t*>(q_buffer),
q_num, q_heads_per_kv, q_num_stride,
q_head_stride, head_dim, inv_scale);
}
// reshape KV to AMX-FP8 friendly layout.
// K: native FP8 layout (1 byte/element, 16-token groups × k-slices).
// V: native VNNI4 packing.
static void reshape_and_cache(
const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
kv_cache_t* __restrict__ key_cache, kv_cache_t* __restrict__ value_cache,
const int64_t* __restrict__ slot_mapping, const int64_t token_num,
const int64_t key_token_num_stride, const int64_t value_token_num_stride,
const int64_t head_num, const int64_t key_head_num_stride,
const int64_t value_head_num_stride, const int64_t num_blocks,
const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
const int64_t block_size, const int64_t /*block_size_stride*/,
const float k_inv = 0.0f, const float v_inv = 0.0f) {
constexpr auto qfn = select_fp8_quant_fn<kv_cache_t>();
// K: native AMX-FP8 layout
reshape_and_cache_k_amx_fp8_impl<scalar_t, qfn>(
key, reinterpret_cast<uint8_t*>(key_cache), slot_mapping, token_num,
head_num, head_dim, block_size, key_token_num_stride,
key_head_num_stride, num_blocks_stride, cache_head_num_stride, k_inv);
{
constexpr int64_t token_num_per_sub_group = 4;
constexpr int64_t head_elems_per_group = 16; // AMX_TILE_ROW_BYTES/4
const int64_t group_num = head_dim / head_elems_per_group;
const int64_t group_size = block_size * head_elems_per_group;
uint8_t* vc = reinterpret_cast<uint8_t*>(value_cache);
#pragma omp parallel for collapse(2) schedule(static)
for (int64_t tok = 0; tok < token_num; ++tok) {
for (int64_t h = 0; h < head_num; ++h) {
const int64_t slot = slot_mapping[tok];
if (slot < 0) continue;
const int64_t block_idx = slot / block_size;
const int64_t block_offset = slot % block_size;
const int64_t sub_group_idx = block_offset / token_num_per_sub_group;
const int64_t sub_group_offset =
block_offset % token_num_per_sub_group;
const scalar_t* vsrc =
value + tok * value_token_num_stride + h * value_head_num_stride;
uint8_t* vdst =
vc + block_idx * num_blocks_stride + h * cache_head_num_stride +
sub_group_idx * token_num_per_sub_group * head_elems_per_group +
sub_group_offset;
for (int64_t i = 0; i < group_num; ++i) {
for (int64_t j = 0; j < head_elems_per_group; ++j)
vdst[j * token_num_per_sub_group] =
qfn(static_cast<float>(vsrc[j]), v_inv);
vsrc += head_elems_per_group;
vdst += group_size;
}
}
}
}
}
private:
alignas(64) __tilecfg amx_tile_config_;
int32_t current_q_head_num_;
};
} // namespace cpu_attention
#endif // CPU_CAPABILITY_AMXFP8
#endif // CPU_ATTN_AMX_FP8_HPP
+54 -36
View File
@@ -12,7 +12,7 @@
#include "cpu/utils.hpp"
namespace cpu_attention {
enum class ISA { AMX, VEC, VEC16, NEON, VXE, RVV, VSX };
enum class ISA { AMX, VEC, VEC16, NEON, VXE, RVV, VSX, AMX_FP8 };
// Mirrors csrc/attention/dtype_fp8.cuh Fp8KVCacheDataType exactly.
enum class Fp8KVCacheDataType {
@@ -152,6 +152,9 @@ struct AttentionMetadata {
case ISA::AMX:
ss << "AMX, ";
break;
case ISA::AMX_FP8:
ss << "AMX_FP8, ";
break;
case ISA::VEC:
ss << "VEC, ";
break;
@@ -885,6 +888,8 @@ struct AttentionInput {
// FP8 KV cache scales (used by FP8 attention implementations)
float k_scale_fp8 = 1.0f;
float v_scale_fp8 = 1.0f;
// FP8 query scale (used by AMX_FP8 which quantizes query inside the kernel)
float q_scale_fp8 = 1.0f;
};
#define DEFINE_CPU_ATTENTION_PARAMS \
@@ -1092,11 +1097,6 @@ class AttentionMainLoop {
// process logits
{
// if (debug_info){
// print_logits("raw logits", logits_buffer, q_head_num,
// kv_tile_token_num, kv_tile_token_num);
// }
if (softcap_scale != 0.0f) {
apply_softcap(logits_buffer, kv_tile_token_num, q_head_num,
kv_tile_token_num, softcap_scale);
@@ -1127,30 +1127,33 @@ class AttentionMainLoop {
apply_softmax(logits_buffer, partial_q_buffer, max_buffer, sum_buffer,
kv_tile_token_num, q_head_num, kv_tile_token_num,
is_first_iter, use_sink);
// if (debug_info){
// print_logits("softmax logits",
// reinterpret_cast<prob_buffer_t*>(logits_buffer), q_head_num,
// kv_tile_token_num, kv_tile_token_num * sizeof(logits_buffer_t) /
// sizeof(prob_buffer_t));
// print_logits("new_max", max_buffer, 1, q_head_num, q_head_num);
// print_logits("new_sum", sum_buffer, 1, q_head_num, q_head_num);
// }
}
// compute P@V
{
constexpr bool prequantize_probabilities = []() {
if constexpr (requires { tile_gemm_t::prequantize_probabilities; }) {
return tile_gemm_t::prequantize_probabilities;
}
return false;
}();
int32_t curr_group_offset =
start_block_group_offset * v_cache_token_group_stride;
int32_t curr_group_num_in_block =
token_group_num_per_block - start_block_group_offset;
int32_t remaining_group_num = token_group_num;
int32_t head_dim_group_num = head_dim / headdim_alignment;
prob_buffer_t* curr_prob_buffer =
reinterpret_cast<prob_buffer_t*>(logits_buffer);
int64_t prob_buffer_stride =
kv_tile_token_num *
(sizeof(logits_buffer_t) / sizeof(prob_buffer_t));
using pv_prob_buffer_t = std::conditional_t<prequantize_probabilities,
uint8_t, prob_buffer_t>;
pv_prob_buffer_t* curr_prob_buffer =
reinterpret_cast<pv_prob_buffer_t*>(logits_buffer);
const int64_t prob_buffer_stride =
prequantize_probabilities
? kv_tile_token_num
: kv_tile_token_num *
(sizeof(logits_buffer_t) / sizeof(prob_buffer_t));
constexpr int64_t prob_buffer_elem_size =
prequantize_probabilities ? sizeof(uint8_t) : sizeof(prob_buffer_t);
partial_output_buffer_t* curr_partial_q_buffer = partial_q_buffer;
bool accum_c = !is_first_iter;
for (int32_t block_idx = start_block_idx; block_idx < end_block_idx;
@@ -1184,15 +1187,13 @@ class AttentionMainLoop {
remaining_group_num -= curr_group_num_in_block;
curr_group_offset = 0;
curr_group_num_in_block = token_group_num_per_block;
curr_prob_buffer += curr_token_num;
curr_prob_buffer = reinterpret_cast<pv_prob_buffer_t*>(
reinterpret_cast<uint8_t*>(curr_prob_buffer) +
curr_token_num * prob_buffer_elem_size);
curr_partial_q_buffer = partial_q_buffer;
accum_c = true;
}
}
// if (debug_info) {
// print_logits("output", partial_q_buffer, q_head_num, head_dim,
// head_dim);
// }
}
void apply_mask(logits_buffer_t* __restrict__ logits_buffer,
@@ -1262,9 +1263,19 @@ class AttentionMainLoop {
#endif
using prob_buffer_vec_t = typename VecTypeTrait<prob_buffer_t>::vec_t;
constexpr bool prequantize_probabilities = []() {
if constexpr (requires { tile_gemm_t::prequantize_probabilities; }) {
return tile_gemm_t::prequantize_probabilities;
}
return false;
}();
using softmax_prob_buffer_t =
std::conditional_t<prequantize_probabilities, uint8_t, prob_buffer_t>;
static_assert(sizeof(prob_buffer_t) <= sizeof(logits_buffer_t));
logits_buffer_t* __restrict__ curr_logits_buffer = logits_buffer;
softmax_prob_buffer_t* __restrict__ curr_prob_buffer =
reinterpret_cast<softmax_prob_buffer_t*>(logits_buffer);
float* __restrict__ curr_partial_q_buffer = partial_q_buffer;
const int32_t vec_num = kv_tile_token_num / 16;
const int32_t head_vec_num = head_dim / 16;
@@ -1302,8 +1313,8 @@ class AttentionMainLoop {
{
logits_buffer_t* __restrict__ curr_logits_buffer_iter =
curr_logits_buffer;
prob_buffer_t* __restrict__ curr_prob_buffer_iter =
reinterpret_cast<prob_buffer_t*>(curr_logits_buffer);
softmax_prob_buffer_t* __restrict__ curr_prob_buffer_iter =
curr_prob_buffer;
for (int32_t j = 0; j < vec_num; ++j) {
vec_op::FP32Vec16 vec(curr_logits_buffer_iter);
vec = vec - max_vec;
@@ -1320,8 +1331,12 @@ class AttentionMainLoop {
vec = fast_exp(vec);
}
prob_buffer_vec_t output_vec(vec);
output_vec.save(curr_prob_buffer_iter);
if constexpr (prequantize_probabilities) {
vec.save(curr_logits_buffer_iter);
} else {
prob_buffer_vec_t output_vec(vec);
output_vec.save(curr_prob_buffer_iter);
}
#else
vec.save(curr_logits_buffer_iter);
for (int32_t k = 0; k < 16; ++k) {
@@ -1335,6 +1350,10 @@ class AttentionMainLoop {
curr_logits_buffer_iter += 16;
curr_prob_buffer_iter += 16;
}
if constexpr (prequantize_probabilities) {
tile_gemm_t::quantize_probability_row(
curr_logits_buffer, curr_prob_buffer, kv_tile_token_num);
}
}
float new_sum_val = sum_vec.reduce_sum();
@@ -1366,6 +1385,12 @@ class AttentionMainLoop {
sum_buffer[i] = new_sum_val;
curr_logits_buffer += logits_buffer_stride;
if constexpr (prequantize_probabilities) {
curr_prob_buffer += kv_tile_token_num;
} else {
curr_prob_buffer =
reinterpret_cast<softmax_prob_buffer_t*>(curr_logits_buffer);
}
curr_partial_q_buffer += head_dim;
}
}
@@ -1840,13 +1865,6 @@ class AttentionMainLoop {
float* curr_sum_buffer = sum_buffer + q_tile_head_offset;
bool debug_info = false;
// bool debug_info = (
// q_head_start_idx == 4 &&
// (q_token_start_idx + q_head_tile_token_offset) <=
// 4
// && (q_token_start_idx + q_head_tile_token_offset +
// q_tile_token_num) > 4
// );
// if (debug_info) {
// std::printf("\tq_iter_idx: %d, q_token_start: %d,"
// "q_token_end: %d, q_token_num: %d, q_head_num: %d,"
+58
View File
@@ -350,6 +350,64 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
};
#endif
// ---------------------------------------------------------------------------
// Vectorized BF16 → FP8 E4M3 quantization helpers.
// ---------------------------------------------------------------------------
// Quantize 16 FP32 values to 16 FP8-E4M3 bytes using AVX-512.
// inv_scale = 1.0f / q_scale (pre-computed by caller).
// Returns 16 packed uint8_t in the low 128-bits of an __m128i.
#if defined(__AVX512F__)
FORCE_INLINE __m128i quant_fp32x16_to_fp8e4m3_avx512(const float* src,
float inv_scale) {
__m512 v = _mm512_loadu_ps(src);
v = _mm512_mul_ps(v, _mm512_set1_ps(inv_scale));
// Clamp to FP8-E4M3 representable range [-448, 448]
v = _mm512_min_ps(v, _mm512_set1_ps(448.0f));
v = _mm512_max_ps(v, _mm512_set1_ps(-448.0f));
// Shift exponent bias: FP32 bias=127 → E4M3 bias=7, delta=120
// Multiply by 2^-120 to move the exponent bias from 127 to 7.
v = _mm512_mul_ps(v, _mm512_set1_ps(0x1p-120f));
__m512i vi = _mm512_castps_si512(v);
// sign: bit31 → bit7
__m512i sign = _mm512_srli_epi32(
_mm512_and_si512(vi, _mm512_set1_epi32(0x80000000u)), 24);
// payload: bits[26:20] -> bits[6:0] after the exponent-bias shift.
__m512i payload = _mm512_srli_epi32(
_mm512_and_si512(vi, _mm512_set1_epi32(0x07F00000u)), 20);
// Keep 0x7F (all-ones) reserved as NaN encoding: clamp to 0x7E
payload = _mm512_min_epu32(payload, _mm512_set1_epi32(0x7E));
__m512i fp8 = _mm512_or_si512(sign, payload);
// Pack 16 × int32 → 16 × uint8 (cvt truncates, so the int32 values must
// fit in uint8; they do since max is 0xFF = sign|payload).
return _mm512_cvtepi32_epi8(fp8);
}
// Quantize 32 BF16 values to 32 FP8-E4M3 bytes using AVX-512.
// Writes 32 bytes to dst. inv_scale = 1.0f / q_scale.
FORCE_INLINE void quant_bf16x32_to_fp8e4m3_avx512(const c10::BFloat16* src,
uint8_t* dst,
float inv_scale) {
// Convert 32 BF16 → 2x16 FP32 then quantize each half.
const uint16_t* u16 = reinterpret_cast<const uint16_t*>(src);
__m256i b16_lo = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(u16));
__m256i b16_hi =
_mm256_loadu_si256(reinterpret_cast<const __m256i*>(u16 + 16));
// Zero-extend uint16 -> uint32, then shift left by 16 to form FP32 bits.
__m512 lo_fp32 =
_mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(b16_lo), 16));
__m512 hi_fp32 =
_mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(b16_hi), 16));
alignas(64) float fp32_buf[32];
_mm512_store_ps(fp32_buf, lo_fp32);
_mm512_store_ps(fp32_buf + 16, hi_fp32);
__m128i lo8 = quant_fp32x16_to_fp8e4m3_avx512(fp32_buf, inv_scale);
__m128i hi8 = quant_fp32x16_to_fp8e4m3_avx512(fp32_buf + 16, inv_scale);
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst), lo8);
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + 16), hi8);
}
#endif // __AVX512F__
struct FP32Vec4 : public Vec<FP32Vec4> {
constexpr static int VEC_ELEM_NUM = 4;
union AliasReg {
+30 -9
View File
@@ -22,6 +22,7 @@ ISA_TYPES = {
"VXE": 4,
"RVV": 5,
"VSX": 6,
"AMX_FP8": 7,
}
# KV cache index: 0 = auto (same as scalar_t), 1 = fp8_e4m3, 2 = fp8_e5m2
@@ -39,13 +40,13 @@ KV_CACHE_CPP_TYPES = {
}
# ISAs supported for head_dims divisible by 32
ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16", "VXE", "RVV", "VSX"]
ISA_FOR_32 = ["AMX", "AMX_FP8", "NEON", "VEC", "VEC16", "VXE", "RVV", "VSX"]
# ISAs supported for head_dims divisible by 16 only
ISA_FOR_16 = ["VEC16"]
# ISAs that support FP8 KV cache (x86 AVX2/AVX-512 required)
ISA_FOR_FP8 = ["AMX", "VEC"]
ISA_FOR_FP8 = ["AMX_FP8", "AMX", "VEC"]
def encode_params(head_dim: int, isa_type: str, kv_cache: str = "auto") -> int:
@@ -64,12 +65,21 @@ def _make_case(
encoded = encode_params(head_dim, isa, kv_cache)
actual_isa = isa_override if isa_override else isa
cpp_type = KV_CACHE_CPP_TYPES[kv_cache]
attn_impl = (
f"cpu_attention::AttentionImpl<"
f"cpu_attention::ISA::{actual_isa}, \\\n"
f" "
native_impl = (
f"cpu_attention::AttentionImpl<cpu_attention::ISA::{actual_isa}, "
f"scalar_t, head_dim, {cpp_type}>"
)
if actual_isa == "AMX_FP8":
fallback_impl = (
f"cpu_attention::AttentionImpl<cpu_attention::ISA::AMX, "
f"scalar_t, head_dim, {cpp_type}>"
)
attn_impl = (
"std::conditional_t<std::is_same_v<scalar_t, c10::BFloat16>, "
f"{native_impl}, {fallback_impl}>"
)
else:
attn_impl = native_impl
comment = (
f"head_dim={head_dim}, isa={isa}"
if kv_cache == "auto"
@@ -98,7 +108,8 @@ def generate_cases_for_isa_group(isa_list: list[str], include_fp8: bool = False)
for isa in isa_list:
if isa not in ISA_FOR_32:
continue
cases.append(_make_case(head_dim, isa, "auto"))
isa_override = "AMX" if isa == "AMX_FP8" else None
cases.append(_make_case(head_dim, isa, "auto", isa_override))
# Non-FP8 cases for head_dims divisible by 16 only
for head_dim in HEAD_DIMS_16:
@@ -112,7 +123,8 @@ def generate_cases_for_isa_group(isa_list: list[str], include_fp8: bool = False)
for isa in isa_list:
if isa not in ISA_FOR_FP8:
continue
cases.append(_make_case(head_dim, isa, fp8_type))
isa_override = "AMX" if isa == "AMX_FP8" and head_dim % 64 else None
cases.append(_make_case(head_dim, isa, fp8_type, isa_override))
return "\n".join(cases)
@@ -138,6 +150,10 @@ def generate_header_file() -> str:
#include "cpu_attn_vec.hpp"
#include "cpu_attn_vec16.hpp"
#ifdef CPU_CAPABILITY_AMXFP8
#include "cpu_attn_amx_fp8.hpp"
#endif
#ifdef CPU_CAPABILITY_AMXBF16
#include "cpu_attn_amx.hpp"
#endif
@@ -206,7 +222,12 @@ def generate_header_file() -> str:
)
header += _macro_block(
"#if defined(CPU_CAPABILITY_AMXBF16)",
"#if defined(CPU_CAPABILITY_AMXFP8)",
["AMX_FP8", "AMX", "VEC", "VEC16"],
fp8=True,
)
header += _macro_block(
"#elif defined(CPU_CAPABILITY_AMXBF16)",
["AMX", "VEC", "VEC16"],
fp8=True,
)