From 30c9d1e47ba5fb9aba270141ab93086aa80ac153 Mon Sep 17 00:00:00 2001 From: Alberto Cabrera Date: Thu, 11 Dec 2025 17:53:29 +0000 Subject: [PATCH] wip: skeleton for q8_0 repack --- ggml/src/ggml-cpu/arch-fallback.h | 22 ++ ggml/src/ggml-cpu/arch/arm/repack.cpp | 102 ++++++++ ggml/src/ggml-cpu/repack.cpp | 347 +++++++++++++++++++++++++- ggml/src/ggml-cpu/repack.h | 8 + 4 files changed, 477 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 0775c87f98..18beff50d3 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -43,6 +43,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 @@ -51,6 +53,8 @@ #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) // repack.cpp #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 @@ -91,6 +95,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 @@ -99,6 +105,8 @@ #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__loongarch64) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -119,6 +127,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 @@ -127,6 +137,8 @@ #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__riscv) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -154,6 +166,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K @@ -161,6 +175,8 @@ #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__s390x__) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -187,6 +203,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 @@ -195,6 +213,8 @@ #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__wasm__) // quants.c #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1 @@ -223,6 +243,8 @@ #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 683ed8d2df..ef234bdf1b 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -2608,3 +2608,105 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, #endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); } + +void ggml_gemv_q8_0_4x4_q8_0(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + // TODO: Implement ARM NEON DOTPROD kernel for q8_0 × q8_0 GEMV +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + ggml_gemv_q8_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_q8_0_4x8_q8_0(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + // TODO: Implement ARM NEON I8MM kernel for q8_0 × q8_0 GEMV +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + ggml_gemv_q8_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_q8_0_4x4_q8_0(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + // TODO: Implement ARM NEON DOTPROD kernel for q8_0 × q8_0 GEMM +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + ggml_gemm_q8_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_q8_0_4x8_q8_0(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + // TODO: Implement ARM NEON I8MM kernel for q8_0 × q8_0 GEMM +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + ggml_gemm_q8_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 9f0d449bd6..c47480f291 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -802,6 +802,219 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, } } +void ggml_gemv_q8_0_4x4_q8_0_generic(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * blocklen + i]; + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j]; + } + } +} + +void ggml_gemv_q8_0_4x8_q8_0_generic(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * blocklen + i]; + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j]; + } + } +} + +void ggml_gemm_q8_0_4x4_q8_0_generic(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + { + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; + } + sumf[m][j] += + sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } + } + } +} + +void ggml_gemm_q8_0_4x8_q8_0_generic(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; + } + sumf[m][j] += + sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } + } +} + void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -1221,6 +1434,23 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs } // extern "C" +static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_interleave) { + block_q8_0x4 out; + + for (int i = 0; i < 4; i++) { + out.d[i] = in[i].d; + } + + const int end = QK8_0 * 4 / blck_size_interleave; + for (int i = 0; i < end; ++i) { + int src_id = i % 4; + int src_offset = (i / 4) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave); + } + return out; +} + static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { block_q4_0x4 out; @@ -1534,6 +1764,38 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block GGML_UNUSED(data_size); } +static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor * t, + int interleave_block, + const void * GGML_RESTRICT data, + size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q8_0); + GGML_ASSERT(interleave_block == 4 || interleave_block == 8); + constexpr int nrows_interleaved = 4; + + block_q8_0x4 * dst = (block_q8_0x4 *) t->data; + const block_q8_0 * src = (const block_q8_0 *) data; + block_q8_0 dst_tmp[4]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK8_0; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q8_0x4(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; +} + static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) { block_iq4_nlx4 out; @@ -1702,6 +1964,14 @@ template <> int repack(struct ggml_tensor * t, const void * return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size); } +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size); +} + // gemv template void gemv(int, float *, size_t, const void *, const void *, int, int); @@ -1738,6 +2008,28 @@ template <> void gemv(int n, float * s, size ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> +void gemv(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { + ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> +void gemv(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { + ggml_gemv_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + // gemm template void gemm(int, float *, size_t, const void *, const void *, int, int); @@ -1758,11 +2050,25 @@ template <> void gemm(int n, float * s, size_t ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> +void gemm(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); } -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> +void gemm(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); } @@ -1774,6 +2080,28 @@ template <> void gemm(int n, float * s, size ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> +void gemm(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { + ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> +void gemm(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { + ggml_gemm_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + class tensor_traits_base : public ggml::cpu::tensor_traits { public: virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; @@ -2168,6 +2496,10 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons static const ggml::cpu::repack::tensor_traits iq4_nl_4x4_q8_0; static const ggml::cpu::repack::tensor_traits iq4_nl_8x8_q8_0; + // instance for Q8_0 + static const ggml::cpu::repack::tensor_traits q8_0_4x4_q8_0; + static const ggml::cpu::repack::tensor_traits q8_0_4x8_q8_0; + if (cur->type == GGML_TYPE_Q4_0) { if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { if (cur->ne[1] % 8 == 0) { @@ -2217,6 +2549,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &iq4_nl_4x4_q8_0; } } + } else if (cur->type == GGML_TYPE_Q8_0) { + if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + if (cur->ne[1] % 4 == 0) { + return &q8_0_4x8_q8_0; + } + } + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + if (cur->ne[1] % 4 == 0) { + return &q8_0_4x4_q8_0; + } + } } return nullptr; diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index c4d928cd15..af98e70344 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -98,6 +98,10 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); // Native implementations void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); @@ -120,6 +124,10 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); #if defined(__cplusplus) } // extern "C"