diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index e4c0fa8d0240c..2b5c6200277c4 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -28,7 +28,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ggml-cpu/repack.h ggml-cpu/hbm.cpp ggml-cpu/hbm.h - ggml-cpu/quants.c ggml-cpu/quants.h ggml-cpu/traits.cpp ggml-cpu/traits.h @@ -453,6 +452,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name) list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC) endif() + # Include generic implementations last to avoid accidentally overriding arch-specific implementations in static linking + list (APPEND GGML_CPU_SOURCES + ggml-cpu/repack-impl.cpp + ggml-cpu/quants.c + ) + if (GGML_CPU_REPACK) target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK) endif() diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 9337e01b62390..2db88ad8d751d 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -14,7 +14,6 @@ #include // for qsort #include // for GGML_ASSERT -#define GGML_CPU_CLANG_WORKAROUND #include "../../repack.h" #if defined(__GNUC__) diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp index 0882b41024362..0cec620b39648 100644 --- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp +++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp @@ -14,7 +14,6 @@ #include // for qsort #include // for GGML_ASSERT -#define GGML_CPU_CLANG_WORKAROUND #include "../../repack.h" #if defined(__GNUC__) diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp index e7635a294a796..42ac9f7c42101 100644 --- a/ggml/src/ggml-cpu/arch/x86/repack.cpp +++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp @@ -14,7 +14,6 @@ #include // for qsort #include // for GGML_ASSERT -#define GGML_CPU_CLANG_WORKAROUND #include "../../repack.h" #if defined(__GNUC__) diff --git a/ggml/src/ggml-cpu/repack-impl.cpp b/ggml/src/ggml-cpu/repack-impl.cpp new file mode 100644 index 0000000000000..069a3a1b942cd --- /dev/null +++ b/ggml/src/ggml-cpu/repack-impl.cpp @@ -0,0 +1,752 @@ +#define GGML_COMMON_IMPL_CPP +#define GGML_COMMON_DECL_CPP +#include "ggml-common.h" +#include "ggml-backend-impl.h" + +#include "ggml-impl.h" +#include "ggml-cpu.h" +#include "ggml-cpu-impl.h" +#include "traits.h" + +#include +#include +#include +#include // for qsort +#include // for GGML_ASSERT + +#include "repack.h" + +#if defined(__GNUC__) +#pragma GCC diagnostic ignored "-Woverlength-strings" +#endif + +#define UNUSED GGML_UNUSED + +static inline int nearest_int(float fval) { + assert(fabsf(fval) <= 4194303.f); + float val = fval + 12582912.f; + int i; memcpy(&i, &val, sizeof(int)); + return (i & 0x007fffff) - 0x00400000; +} + +// Functions to create the interleaved data layout formats + +// interleave 4 block_q4_0s in blocks of blck_size_interleave +// returns an interleaved block_q4_0x4 +// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks +// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave +// +// - in : an array of block_q4_0 pointers +// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of +// blck_size_interleave bytes +// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes +// from bias offset form to pure sign form (this saves subtract +// operations durin unpacking) +// + +void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; + + // scalar + const int blck_size_interleave = 4; + float srcv[4][QK8_0]; + float id[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; + amax = MAX(amax, fabsf(srcv[row_iter][j])); + } + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < QK8_0 * 4; j++) { + int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; + int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; + src_offset += (j % blck_size_interleave); + + float x0 = srcv[src_id][src_offset] * id[src_id]; + y[i].qs[j] = roundf(x0); + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x4) + +void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; + + // scalar + const int blck_size_interleave = 8; + float srcv[4][QK8_0]; + float id[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; + amax = MAX(amax, fabsf(srcv[row_iter][j])); + } + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < QK8_0 * 4; j++) { + int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; + int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; + src_offset += (j % blck_size_interleave); + + float x0 = srcv[src_id][src_offset] * id[src_id]; + y[i].qs[j] = roundf(x0); + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x8) + +void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK_K == 256); + assert(k % QK_K == 0); + const int nb = k / QK_K; + + block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy; + + // scalar + const int blck_size_interleave = 8; + float srcv[4][QK_K]; + float iscale[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + float max = 0; + + for (int j = 0; j < QK_K; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK_K + j]; + // Update the maximum value of the corresponding super block + if(amax < fabsf(srcv[row_iter][j])) { + amax = fabsf(srcv[row_iter][j]); + max = srcv[row_iter][j]; + } + } + + iscale[row_iter] = amax ? -127.f/max : 0; + + y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; + } + + for (int j = 0; j < QK_K / 4; j++) { + y[i].bsums[j] = 0; + } + + // Quants values are interleaved in sequence of eight bytes from corresponding super blocks + // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving + // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on + for (int j = 0; j < QK_K * 4; j++) { + int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; + int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; + src_offset += (j % blck_size_interleave); + int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); + + float x0 = srcv[src_id][src_offset] * iscale[src_id]; + y[i].qs[j] = nearest_int(x0); + y[i].bsums[index] += y[i].qs[j]; + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_K_4x8) + +void ggml_gemv_q4_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 (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x4_q8_0) + +void ggml_gemv_q4_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(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x8_q8_0) + +void ggml_gemv_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; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + { + float sumf[8]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) 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 / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_8x8_q8_0) + +void ggml_gemv_q4_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) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + static const uint32_t kmask1 = 0x3f3f3f3f; + static const uint32_t kmask2 = 0x0f0f0f0f; + static const uint32_t kmask3 = 0x03030303; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[8]; + float sum_minf[8]; + uint32_t utmp[32]; + int sumi1; + int sumi2; + int sumi; + + const block_q8_K * a_ptr = (const block_q8_K *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0; + sum_minf[j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int sb = 0; sb < 8; sb++) { + memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); + utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; + utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); + utmp[sb * 4 + 2] = uaux_0; + utmp[sb * 4 + 0] &= kmask1; + } + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; + uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); + sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]); + sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + } + } + for (int sb = 0; sb < 8; sb++) { + uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; + for (int j = 0; j < ncols_interleaved; j++) { + sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_K_8x8_q8_K) + +void ggml_gemv_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) { + 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(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) 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 / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemv_iq4_nl_4x4_q8_0) + +void ggml_gemm_q4_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(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; + } + } + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x4_q8_0) + +void ggml_gemm_q4_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(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; + } + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x8_q8_0) + +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; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4][8]; + 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_q4_0x8 * b_ptr = (const block_q4_0x8 *) 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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; + } + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_8x8_q8_0) + +void ggml_gemm_q4_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) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + static const uint32_t kmask1 = 0x3f3f3f3f; + static const uint32_t kmask2 = 0x0f0f0f0f; + static const uint32_t kmask3 = 0x03030303; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4][8]; + float sum_minf[4][8]; + uint32_t utmp[32]; + int sumi1; + int sumi2; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + sum_minf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int sb = 0; sb < 8; sb++) { + memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); + utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; + utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); + utmp[sb * 4 + 2] = uaux_0; + utmp[sb * 4 + 0] &= kmask1; + } + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; + uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); + sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]); + sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; + } + } + } + for (int sb = 0; sb < 8; sb++) { + uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; + for(int m = 0; m < 4; m++) { + const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); + for(int j = 0; j < ncols_interleaved; j++) { + sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * 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] - sum_minf[m][j]; + } + } + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_K_8x8_q8_K) + +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) { + 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(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) 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 / (2 * 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 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; + } + } + } + } +} +GGML_CPU_NATIVE_IMPL(ggml_gemm_iq4_nl_4x4_q8_0) diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 604ccee907843..20e7baebb72e0 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1,5 +1,3 @@ -#define GGML_COMMON_IMPL_CPP -#define GGML_COMMON_DECL_CPP #include "ggml-common.h" #include "ggml-backend-impl.h" @@ -26,162 +24,6 @@ #define UNUSED GGML_UNUSED -static inline int nearest_int(float fval) { - assert(fabsf(fval) <= 4194303.f); - float val = fval + 12582912.f; - int i; memcpy(&i, &val, sizeof(int)); - return (i & 0x007fffff) - 0x00400000; -} - -// Functions to create the interleaved data layout formats - -// interleave 4 block_q4_0s in blocks of blck_size_interleave -// returns an interleaved block_q4_0x4 -// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks -// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave -// -// - in : an array of block_q4_0 pointers -// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of -// blck_size_interleave bytes -// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes -// from bias offset form to pure sign form (this saves subtract -// operations durin unpacking) -// - -extern "C" { - -void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { - assert(QK8_0 == 32); - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; - - // scalar - const int blck_size_interleave = 4; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x4) - -void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { - assert(QK8_0 == 32); - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; - - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x8) - -void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { - assert(QK_K == 256); - assert(k % QK_K == 0); - const int nb = k / QK_K; - - block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy; - - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK_K]; - float iscale[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - float max = 0; - - for (int j = 0; j < QK_K; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK_K + j]; - // Update the maximum value of the corresponding super block - if(amax < fabsf(srcv[row_iter][j])) { - amax = fabsf(srcv[row_iter][j]); - max = srcv[row_iter][j]; - } - } - - iscale[row_iter] = amax ? -127.f/max : 0; - - y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; - } - - for (int j = 0; j < QK_K / 4; j++) { - y[i].bsums[j] = 0; - } - - // Quants values are interleaved in sequence of eight bytes from corresponding super blocks - // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving - // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on - for (int j = 0; j < QK_K * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); - - float x0 = srcv[src_id][src_offset] * iscale[src_id]; - y[i].qs[j] = nearest_int(x0); - y[i].bsums[index] += y[i].qs[j]; - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_K_4x8) - -} // extern "C" - template void ggml_quantize_mat_t(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row); @@ -203,587 +45,6 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row); } -extern "C" { - -void ggml_gemv_q4_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 (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x4_q8_0) - -void ggml_gemv_q4_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(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x8_q8_0) - -void ggml_gemv_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; - const int ncols_interleaved = 8; - const int blocklen = 8; - - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - - { - float sumf[8]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) 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 / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_8x8_q8_0) - -void ggml_gemv_q4_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) { - const int qk = QK_K; - const int nb = n / qk; - const int ncols_interleaved = 8; - const int blocklen = 8; - static const uint32_t kmask1 = 0x3f3f3f3f; - static const uint32_t kmask2 = 0x0f0f0f0f; - static const uint32_t kmask3 = 0x03030303; - - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - - float sumf[8]; - float sum_minf[8]; - uint32_t utmp[32]; - int sumi1; - int sumi2; - int sumi; - - const block_q8_K * a_ptr = (const block_q8_K *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) { - sumf[j] = 0.0; - sum_minf[j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int sb = 0; sb < 8; sb++) { - memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); - utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); - const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; - utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); - utmp[sb * 4 + 2] = uaux_0; - utmp[sb * 4 + 0] &= kmask1; - } - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; - uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; - for (int j = 0; j < ncols_interleaved; j++) { - sumi1 = 0; - sumi2 = 0; - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); - sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]); - sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]); - sumi1 = sumi1 * scales_0[j]; - sumi2 = sumi2 * scales_1[j]; - sumi += sumi1 + sumi2; - } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; - } - } - for (int sb = 0; sb < 8; sb++) { - uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; - for (int j = 0; j < ncols_interleaved; j++) { - sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; - } - } - } - for (int j = 0; j < ncols_interleaved; j++) { - s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_K_8x8_q8_K) - -void ggml_gemv_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) { - 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(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) 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 / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; - const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); - } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemv_iq4_nl_4x4_q8_0) - -void ggml_gemm_q4_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(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; - } - } - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x4_q8_0) - -void ggml_gemm_q4_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(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_q4_0x4 * b_ptr = (const block_q4_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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; - } - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x8_q8_0) - -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; - const int ncols_interleaved = 8; - const int blocklen = 8; - - assert (n % qk == 0); - assert (nr % 4 == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - - float sumf[4][8]; - 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_q4_0x8 * b_ptr = (const block_q4_0x8 *) 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 / (2 * 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 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; - } - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_8x8_q8_0) - -void ggml_gemm_q4_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) { - const int qk = QK_K; - const int nb = n / qk; - const int ncols_interleaved = 8; - const int blocklen = 8; - static const uint32_t kmask1 = 0x3f3f3f3f; - static const uint32_t kmask2 = 0x0f0f0f0f; - static const uint32_t kmask3 = 0x03030303; - - assert (n % qk == 0); - assert (nr % 4 == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - - float sumf[4][8]; - float sum_minf[4][8]; - uint32_t utmp[32]; - int sumi1; - int sumi2; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumf[m][j] = 0.0; - sum_minf[m][j] = 0.0; - } - } - for (int l = 0; l < nb; l++) { - for (int sb = 0; sb < 8; sb++) { - memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); - utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); - const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; - utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); - utmp[sb * 4 + 2] = uaux_0; - utmp[sb * 4 + 0] &= kmask1; - } - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; - uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi1 = 0; - sumi2 = 0; - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); - sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]); - sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); - sumi1 = sumi1 * scales_0[j]; - sumi2 = sumi2 * scales_1[j]; - sumi += sumi1 + sumi2; - } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; - } - } - } - for (int sb = 0; sb < 8; sb++) { - uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; - for(int m = 0; m < 4; m++) { - const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); - for(int j = 0; j < ncols_interleaved; j++) { - sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * 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] - sum_minf[m][j]; - } - } - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_K_8x8_q8_K) - -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) { - 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(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) 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 / (2 * 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 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; - const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); - } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_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]; - } - } - } - } -} -GGML_CPU_NATIVE_IMPL(ggml_gemm_iq4_nl_4x4_q8_0) - -} // extern "C" - static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { block_q4_0x4 out; diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index b13d2d0c73495..302b2ec42cd6e 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -60,14 +60,11 @@ struct block_iq4_nlx4 { static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding"); -#if defined(__cplusplus) +// Clang always mangles the alias name when compiling a C++ source file +#if defined(__cplusplus) && !(defined(__GNUC__) && defined(__clang__)) extern "C" { #endif -// Workaround for clang: -// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous'' -// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call) -#if defined(GGML_CPU_CLANG_WORKAROUND) || defined(__APPLE__) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__) void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); @@ -81,7 +78,12 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemm_q4_0_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_gemm_q4_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); -#endif // !defined(__clang__) + +#if defined(__cplusplus) && !(defined(__GNUC__) && defined(__clang__)) +} // extern "C" +#endif + +extern "C" { // Native implementations void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); @@ -98,6 +100,4 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemm_q4_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); -#if defined(__cplusplus) -} // extern "C" -#endif +}