mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-11 17:44:38 +00:00
Merge branch 'master' into concedo_experimental
# Conflicts: # CMakeLists.txt # Makefile # ggml-opencl.c
This commit is contained in:
Concedo
commit
3de34ee492
11 changed files with 260 additions and 95 deletions
189
ggml.c
189
ggml.c
|
|
@ -331,7 +331,7 @@ static ggml_fp16_t table_exp_f16[1 << 16];
|
|||
// precomputed f32 table for f16 (256 KB)
|
||||
static float table_f32_f16[1 << 16];
|
||||
|
||||
#if defined(__ARM_NEON)
|
||||
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
|
||||
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
|
||||
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
|
||||
#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
|
||||
|
|
@ -1096,7 +1096,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
|
|||
const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id));
|
||||
const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f));
|
||||
const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf);
|
||||
const v128_t vc = wasm_i32x4_min_u(vi, wasm_i32x4_splat(15));
|
||||
const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15));
|
||||
|
||||
y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4);
|
||||
y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4);
|
||||
|
|
@ -1993,8 +1993,8 @@ static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, in
|
|||
const uint8_t vi = pp[l/2];
|
||||
|
||||
// extract the 5-th bit from qh
|
||||
const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
|
||||
const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
|
||||
|
||||
const int8_t vi0 = (vi & 0x0F) | vh0;
|
||||
const int8_t vi1 = (vi >> 4) | vh1;
|
||||
|
|
@ -2030,8 +2030,8 @@ static void dequantize_row_q5_1(const void * restrict vx, float * restrict y, in
|
|||
const uint8_t vi = pp[l/2];
|
||||
|
||||
// extract the 5-th bit from qh
|
||||
const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
|
||||
const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
|
||||
|
||||
const uint8_t vi0 = (vi & 0x0F) | vh0;
|
||||
const uint8_t vi1 = (vi >> 4) | vh1;
|
||||
|
|
@ -3401,6 +3401,72 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
|
|||
}
|
||||
|
||||
*s = vaddvq_f32(sumv);
|
||||
#elif defined(__wasm_simd128__)
|
||||
v128_t sumv = wasm_f32x4_splat(0.0f);
|
||||
|
||||
uint64_t tmp[4];
|
||||
|
||||
for (int i = 0; i < nb; ++i) {
|
||||
const block_q5_0 * restrict x0 = &x[i];
|
||||
const block_q8_0 * restrict y0 = &y[i];
|
||||
|
||||
const v128_t m4b = wasm_i8x16_splat(0x0F);
|
||||
const v128_t s16b = wasm_i8x16_splat(0x10);
|
||||
|
||||
// extract the 5th bit
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x0->qh, sizeof(qh));
|
||||
|
||||
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
|
||||
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
|
||||
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
|
||||
tmp[3] = table_b2b_u[(qh >> 24) ];
|
||||
|
||||
const v128_t qhl = wasm_v128_load(tmp + 0);
|
||||
const v128_t qhh = wasm_v128_load(tmp + 2);
|
||||
|
||||
const v128_t v0 = wasm_v128_load(x0->qs);
|
||||
|
||||
// 4-bit -> 8-bit
|
||||
const v128_t v0l = wasm_v128_and (v0, m4b);
|
||||
const v128_t v0h = wasm_u8x16_shr(v0, 4);
|
||||
|
||||
// interleave
|
||||
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
|
||||
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
|
||||
|
||||
// add high bit and sub 16
|
||||
const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b);
|
||||
const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b);
|
||||
|
||||
// load y
|
||||
const v128_t v1l = wasm_v128_load(y0->qs);
|
||||
const v128_t v1h = wasm_v128_load(y0->qs + 16);
|
||||
|
||||
// int8x16 -> int16x8
|
||||
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
|
||||
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
|
||||
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
|
||||
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
|
||||
|
||||
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
|
||||
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
|
||||
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
|
||||
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
|
||||
|
||||
const float x0d = GGML_FP16_TO_FP32(x0->d);
|
||||
|
||||
// dot product
|
||||
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
|
||||
wasm_i32x4_add(
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
|
||||
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
|
||||
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
|
||||
}
|
||||
|
||||
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
|
||||
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
|
||||
#elif defined(__AVX2__)
|
||||
// Initialize accumulator with zeros
|
||||
__m256 acc = _mm256_setzero_ps();
|
||||
|
|
@ -3441,8 +3507,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
|
|||
for (int j = 0; j < QK8_0/2; j++) {
|
||||
const uint8_t v0 = x0[j];
|
||||
|
||||
const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
|
||||
const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
|
||||
const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
|
||||
const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
|
||||
|
||||
const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16;
|
||||
const int x1_0 = ((v0 >> 4) | x1_0h) - 16;
|
||||
|
|
@ -3532,6 +3598,77 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
|
|||
}
|
||||
|
||||
*s = vaddvq_f32(sumv) + summs;
|
||||
#elif defined(__wasm_simd128__)
|
||||
v128_t sumv = wasm_f32x4_splat(0.0f);
|
||||
|
||||
float summs = 0.0f;
|
||||
|
||||
uint64_t tmp[4];
|
||||
|
||||
for (int i = 0; i < nb; ++i) {
|
||||
const block_q5_1 * restrict x0 = &x[i];
|
||||
const block_q8_1 * restrict y0 = &y[i];
|
||||
|
||||
summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1);
|
||||
|
||||
const v128_t m4b = wasm_i8x16_splat(0x0F);
|
||||
|
||||
// extract the 5th bit
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x0->qh, sizeof(qh));
|
||||
|
||||
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
|
||||
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
|
||||
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
|
||||
tmp[3] = table_b2b_u[(qh >> 24) ];
|
||||
|
||||
const v128_t qhl = wasm_v128_load(tmp + 0);
|
||||
const v128_t qhh = wasm_v128_load(tmp + 2);
|
||||
|
||||
const v128_t v0 = wasm_v128_load(x0->qs);
|
||||
|
||||
// 4-bit -> 8-bit
|
||||
const v128_t v0l = wasm_v128_and (v0, m4b);
|
||||
const v128_t v0h = wasm_u8x16_shr(v0, 4);
|
||||
|
||||
static bool x = true;
|
||||
|
||||
// interleave
|
||||
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
|
||||
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
|
||||
|
||||
// add high bit
|
||||
const v128_t v0lf = wasm_v128_or(v0lz, qhl);
|
||||
const v128_t v0hf = wasm_v128_or(v0hz, qhh);
|
||||
|
||||
// load y
|
||||
const v128_t v1l = wasm_v128_load(y0->qs);
|
||||
const v128_t v1h = wasm_v128_load(y0->qs + 16);
|
||||
|
||||
// int8x16 -> int16x8
|
||||
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
|
||||
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
|
||||
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
|
||||
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
|
||||
|
||||
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
|
||||
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
|
||||
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
|
||||
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
|
||||
|
||||
const float x0d = GGML_FP16_TO_FP32(x0->d);
|
||||
|
||||
// dot product
|
||||
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
|
||||
wasm_i32x4_add(
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
|
||||
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
|
||||
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
|
||||
}
|
||||
|
||||
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
|
||||
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
|
||||
#elif defined(__AVX2__)
|
||||
// Initialize accumulator with zeros
|
||||
__m256 acc = _mm256_setzero_ps();
|
||||
|
|
@ -3575,8 +3712,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
|
|||
for (int j = 0; j < QK8_1/2; j++) {
|
||||
const uint8_t v0 = x0[j];
|
||||
|
||||
const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
|
||||
const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
|
||||
const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
|
||||
const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
|
||||
|
||||
const int x0_0 = (v0 & 0x0F) | x0_0h;
|
||||
const int x1_0 = (v0 >> 4) | x1_0h;
|
||||
|
|
@ -4052,6 +4189,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
|
|||
"DIAG_MASK_INF",
|
||||
"SOFT_MAX",
|
||||
"ROPE",
|
||||
"ALIBI",
|
||||
"CONV_1D_1S",
|
||||
"CONV_1D_2S",
|
||||
|
||||
|
|
@ -4100,6 +4238,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
|
|||
"diag_mask_inf(x)",
|
||||
"soft_max(x)",
|
||||
"rope(x)",
|
||||
"alibi(x)",
|
||||
"conv_1d_1s(x)",
|
||||
"conv_1d_2s(x)",
|
||||
|
||||
|
|
@ -4280,6 +4419,28 @@ bool ggml_is_quantized(enum ggml_type type) {
|
|||
return GGML_IS_QUANTIZED[type];
|
||||
}
|
||||
|
||||
enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
|
||||
enum ggml_type wtype = GGML_TYPE_COUNT;
|
||||
|
||||
switch (ftype) {
|
||||
case GGML_FTYPE_ALL_F32: wtype = GGML_TYPE_F32; break;
|
||||
case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_2: wtype = GGML_TYPE_Q4_2; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_3: wtype = GGML_TYPE_Q4_3; break;
|
||||
case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break;
|
||||
case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break;
|
||||
case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break;
|
||||
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
|
||||
}
|
||||
|
||||
GGML_ASSERT(wtype != GGML_TYPE_COUNT);
|
||||
|
||||
return wtype;
|
||||
}
|
||||
|
||||
static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
|
||||
return tensor->nb[0] > tensor->nb[1];
|
||||
}
|
||||
|
|
@ -13149,8 +13310,8 @@ size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t *
|
|||
memcpy(&qh, &y[i].qh, sizeof(qh));
|
||||
|
||||
for (int l = 0; l < QK5_0; l += 2) {
|
||||
const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
|
||||
const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
|
||||
|
||||
// cast to 16 bins
|
||||
const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;
|
||||
|
|
@ -13179,8 +13340,8 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t *
|
|||
memcpy(&qh, &y[i].qh, sizeof(qh));
|
||||
|
||||
for (int l = 0; l < QK5_1; l += 2) {
|
||||
const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
|
||||
const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
|
||||
const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
|
||||
|
||||
// cast to 16 bins
|
||||
const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue