Skip to content

Commit 7dad717

Browse files
pubbysw
pubby
authored and
sw
committed
Faster q3_0 implementation, using two planes, by @pubby 
1 parent e2f8c1e commit 7dad717

File tree

1 file changed

+102
-112
lines changed

1 file changed

+102
-112
lines changed

‎ggml.c‎

Lines changed: 102 additions & 112 deletions
Original file line numberDiff line numberDiff line change
@@ -606,12 +606,12 @@ typedef struct {
606606
static_assert(sizeof(block_q2_0) == sizeof(ggml_fp16_t) + QK2_0 / 4, "wrong q2_0 size/padding");
607607

608608
#define QK3_0 16
609-
typedef union {
610-
struct {
611-
uint16_t pad[3];
612-
ggml_fp16_t d;
613-
};
614-
uint64_t qs;
609+
typedef struct {
610+
ggml_fp16_t d;
611+
// Instead of representing q3_0 as a packed format "...210210210210",
612+
// represent it as two planes: "...10101010" and "...2222"
613+
uint16_t qhi; // The highest bit of each 3-bit number, packed together
614+
uint32_t qlo; // The low 2-bits of each 3-bit number, packed together
615615
} block_q3_0;
616616
static_assert(sizeof(block_q3_0) == sizeof(ggml_fp16_t) + QK3_0 * 3 / 8, "wrong q3_0 size/padding");
617617

@@ -691,17 +691,20 @@ static void quantize_row_q3_0(const float * restrict x, block_q3_0 * restrict y,
691691
const float d = max / -4;
692692
const float id = d ? 1.0f/d : 0.0f;
693693

694-
uint64_t qs = 0;
694+
uint32_t lo = 0;
695+
uint16_t hi = 0;
695696

696697
for (int l = 0; l < QK3_0; l++) {
697698
const float v = x[i*QK3_0 + l]*id;
698699
const uint8_t vi = MIN(7, (int8_t)roundf(v) + 4);
699700
assert(vi < 8);
700-
qs |= (uint64_t)vi << (l*3);
701+
lo |= (vi & 3) << (l * 2);
702+
hi |= ((vi >> 2) & 1) << l;
701703
}
702704

703-
y[i].qs = qs;
704-
y[i].d = GGML_FP32_TO_FP16(d); // overwrite unused part of uint64_t qs
705+
y[i].d = GGML_FP32_TO_FP16(d);
706+
y[i].qlo = lo;
707+
y[i].qhi = hi;
705708
}
706709
}
707710

@@ -1335,13 +1338,15 @@ static void dequantize_row_q3_0(const void * restrict vx, float * restrict y, in
13351338

13361339
for (int i = 0; i < nb; i++) {
13371340
const float d = GGML_FP16_TO_FP32(x[i].d);
1338-
uint64_t qs = x[i].qs;
1341+
uint_fast32_t lo = x[i].qlo;
1342+
uint_fast32_t hi = x[i].qhi << 2;
13391343
for (int l = 0; l < QK3_0; l++) {
1340-
const int8_t vi = qs & 7;
1344+
const int8_t vi = (lo & 3) | (hi & 4);
13411345
const float v = (vi - 4)*d;
13421346
y[i*QK3_0 + l] = v;
13431347
assert(!isnan(y[i*QK3_0 + l]));
1344-
qs >>= 3;
1348+
lo >>= 2;
1349+
hi >>= 1;
13451350
}
13461351
}
13471352
}
@@ -2391,6 +2396,39 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t
23912396
*s = sumf;
23922397
}
23932398

2399+
#if __AVX2__ || __AVX512F__
2400+
// Computes the dot product of signed 8-bit integers packed into 256-bit vectors,
2401+
// converting the result to 32-bit floats packed into a 256-bit vector.
2402+
static inline __m256 dotMul(__m256i bx, __m256i by) {
2403+
# if __AVXVNNIINT8__
2404+
// Perform multiplication and sum to 32-bit values
2405+
const __m256i i32 = _mm256_dpbssd_epi32(bx, by, _mm256_setzero_si256());
2406+
# else
2407+
// Get absolute values of x vectors
2408+
const __m256i ax = _mm256_sign_epi8(bx, bx);
2409+
// Sign the values of the y vectors
2410+
const __m256i sy = _mm256_sign_epi8(by, bx);
2411+
// Perform multiplication and create 16-bit values
2412+
const __m256i dot = _mm256_maddubs_epi16(ax, sy);
2413+
2414+
// Convert int16_t to int32_t by adding pairwise
2415+
const __m256i ones = _mm256_set1_epi16(1);
2416+
const __m256i i32 = _mm256_madd_epi16(ones, dot);
2417+
# endif
2418+
// Convert int32_t to float
2419+
return _mm256_cvtepi32_ps(i32);
2420+
}
2421+
2422+
// Return horizontal sum of 32-bit floats packed into a 256-bit vector.
2423+
static inline float horizontalSum(__m256 acc) {
2424+
__m128 res = _mm256_extractf128_ps(acc, 1);
2425+
res = _mm_add_ps(res, _mm256_castps256_ps128(acc));
2426+
res = _mm_add_ps(res, _mm_movehl_ps(res, res));
2427+
res = _mm_add_ss(res, _mm_movehdup_ps(res));
2428+
return _mm_cvtss_f32(res);
2429+
}
2430+
#endif
2431+
23942432
static void ggml_vec_dot_q2_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
23952433
assert(n % QK2_0 == 0);
23962434
const int nb = n / QK2_0;
@@ -2420,30 +2458,15 @@ static void ggml_vec_dot_q2_0_q8_0(const int n, float * restrict s, const void *
24202458
// Load y vector
24212459
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
24222460

2423-
// Get absolute values of x vectors
2424-
const __m256i ax = _mm256_sign_epi8(bx, bx);
2425-
// Sign the values of the y vectors
2426-
const __m256i sy = _mm256_sign_epi8(by, bx);
2427-
// Perform multiplication and create 16-bit values
2428-
const __m256i dot = _mm256_maddubs_epi16(ax, sy);
2429-
2430-
// Convert int16_t to int32_t by adding pairwise
2431-
const __m256i ones = _mm256_set1_epi16(1);
2432-
__m256i i32 = _mm256_madd_epi16(ones, dot);
2433-
2434-
// Convert int32_t to float
2435-
__m256 p = _mm256_cvtepi32_ps(i32);
2461+
// Do the product:
2462+
__m256 p = dotMul(bx, by);
24362463

24372464
// Apply the scale, and accumulate
24382465
acc = _mm256_fmadd_ps(scale, p, acc);
24392466
}
24402467

24412468
// Return horizontal sum of the acc vector
2442-
__m128 res = _mm256_extractf128_ps(acc, 1);
2443-
res = _mm_add_ps(res, _mm256_castps256_ps128(acc));
2444-
res = _mm_add_ps(res, _mm_movehl_ps(res, res));
2445-
res = _mm_add_ss(res, _mm_movehdup_ps(res));
2446-
sumf = _mm_cvtss_f32(res);
2469+
sumf = horizontalSum(acc);
24472470
#else
24482471
for (int i = 0; i < nb; i++) {
24492472
const float d0 = GGML_FP16_TO_FP32(x[i].d);
@@ -2468,6 +2491,20 @@ static void ggml_vec_dot_q2_0_q8_0(const int n, float * restrict s, const void *
24682491
*s = sumf;
24692492
}
24702493

2494+
// Lookup table used to convert q3_0 to SIMD vectors.
2495+
// Expands the bits of an 8-bit value into a 64 bit result, turning each bit into a byte.
2496+
// A zero bit turns into 0xFC, while a one bit turns into 0x00.
2497+
#define B0(n) 0x ## n
2498+
#define B1(n) B0(n ## FC), B0(n ## 00)
2499+
#define B2(n) B1(n ## FC), B1(n ## 00)
2500+
#define B3(n) B2(n ## FC), B2(n ## 00)
2501+
#define B4(n) B3(n ## FC), B3(n ## 00)
2502+
#define B5(n) B4(n ## FC), B4(n ## 00)
2503+
#define B6(n) B5(n ## FC), B5(n ## 00)
2504+
#define B7(n) B6(n ## FC), B6(n ## 00)
2505+
#define B8( ) B7( FC), B7( 00)
2506+
static const uint64_t ggml_q3_table[256] = { B8() };
2507+
24712508
static void ggml_vec_dot_q3_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
24722509
assert(n % QK3_0 == 0);
24732510
const int nb = n / QK3_0;
@@ -2480,103 +2517,54 @@ static void ggml_vec_dot_q3_0_q8_0(const int n, float * restrict s, const void *
24802517

24812518
#if defined(__AVX2__)
24822519
// Initialize accumulator with zeros
2483-
__m128 acc = _mm_setzero_ps();
2520+
__m256 acc = _mm256_setzero_ps();
2521+
24842522
for (int i = 0; i < nb/2; i++) {
2485-
const __m128 scale_y = _mm_set1_ps(y[i].d);
2486-
for (int u = 0; u < 2; u++) { // let the compiler unroll this
2487-
// Compute combined scale for the block
2488-
const __m128 scale_x = _mm_set1_ps(GGML_FP16_TO_FP32(x[i*2+u].d));
2489-
const __m128 scale = _mm_mul_ps(scale_x, scale_y);
2490-
2491-
__m256i bxx = _mm256_set1_epi64x(x[i*2+u].qs);
2492-
2493-
// legend: _=zero +=one .=don't care 0-f=3bit quantized values s=fp16 scale
2494-
2495-
// shift the copies to be able to reach all values
2496-
// 255 192 128 64 0
2497-
// | | | |
2498-
// sssssfedcba9876543210sssssfedcba9876543210sssssfedcba9876543210sssssfedcba9876543210 in
2499-
// sssfedcba9876543210_______________________sfedcba9876543210____sssssfedcba9876543210 shift left
2500-
// _______________________sssssfedcba98765432__________________________________________ shift right
2501-
// sssfedcba9876543210____sssssfedcba98765432sfedcba9876543210____sssssfedcba9876543210 out
2502-
// ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^
2503-
// e b 6 3 _ . f a 7 2 c 9 4 1 _ . d 8 5 0
2504-
const __m256i shift_l = _mm256_set_epi64x(2*3, 64, 4*3, 0);
2505-
const __m256i shift_r = _mm256_set_epi64x( 64, 2*3, 64, 64);
2506-
bxx = _mm256_or_si256(_mm256_sllv_epi64(bxx, shift_l), _mm256_srlv_epi64(bxx, shift_r));
2507-
2508-
// add to itself in masked places to shift some values left one bit
2509-
// 127 64 0
2510-
// | | | | | | | | | | | | | | | |
2511-
// ssssfffeeedddcccbbbaaa999888777666555444333222111000____________ssssssssssssssssfffeeedddcccbbbaaa999888777666555444333222111000 in
2512-
// _____________________++++____________________++++____________________________________++++____________________++++_______________ mask
2513-
// _____________________.999____________________.111____________________________________.ddd____________________.555_______________ masked
2514-
// .............ccc.....999.............444.....111....____________.....................ddd.............888.....555.............000 sum
2515-
//
2516-
// 255 192 128
2517-
// | | | | | | | | | | | | | | | |
2518-
// ssssssssssfffeeedddcccbbbaaa999888777666555444333222111000____________ssssssssssssssssfffeeedddcccbbbaaa999888777666555444333222 in
2519-
// _____________________++++____________________++++____________________________________++++____________________++++_______________ mask
2520-
// _____________________.bbb____________________.333____________________________________.fff____________________.777_______________ masked
2521-
// .............eee.....bbb.............666.....333..........____________...............fff.............aaa.....777.............222 sum
2522-
const __m256i doublemask = _mm256_set1_epi64x(0x078000078000);
2523-
bxx = _mm256_add_epi64(bxx, _mm256_and_si256(doublemask, bxx));
2524-
2525-
// collect 16 bytes from 256 into 128 bits
2526-
const __m256i shufmask = _mm256_set_epi8(
2527-
5,14,-1,-1,13, 3,-1,-1, 2,11,-1,-1,10, 0,-1,-1,
2528-
-1,-1, 5,14,-1,-1,13, 3,-1,-1, 2,11,-1,-1,10, 0);
2529-
bxx = _mm256_shuffle_epi8(bxx, shufmask);
2530-
2531-
__m128i bx = _mm_or_si128(_mm256_castsi256_si128(bxx), _mm256_extracti128_si256(bxx, 1));
2532-
2533-
const __m128i mask = _mm_set1_epi8(7);
2534-
bx = _mm_and_si128(mask, bx);
2535-
2536-
const __m128i off = _mm_set1_epi8(4);
2537-
bx = _mm_sub_epi8(bx, off);
2538-
2539-
const __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + u*QK3_0));
2523+
__m256i bx = bytesFromCrumbs(x[i*2+1].qlo, x[i*2].qlo);
25402524

2541-
// Get absolute values of x vectors
2542-
const __m128i ax = _mm_sign_epi8(bx, bx);
2543-
// Sign the values of the y vectors
2544-
const __m128i sy = _mm_sign_epi8(by, bx);
2545-
// Perform multiplication and create 16-bit values
2546-
const __m128i dot = _mm_maddubs_epi16(ax, sy);
2525+
__m256i const bxhi = _mm256_set_epi64x(
2526+
ggml_q3_table[x[i*2+1].qhi >> 8], ggml_q3_table[x[i*2+1].qhi & 0xFF],
2527+
ggml_q3_table[x[i*2+0].qhi >> 8], ggml_q3_table[x[i*2+0].qhi & 0xFF]);
25472528

2548-
// Convert int16_t to int32_t by adding pairwise
2549-
const __m128i ones = _mm_set1_epi16(1);
2550-
__m128i i32 = _mm_madd_epi16(dot, ones);
2529+
// OR the high bits (which also handles the sign):
2530+
bx = _mm256_or_si256(bx, bxhi);
2531+
2532+
// Compute combined scale for the block
2533+
const __m128 scale_lo = _mm_set1_ps(GGML_FP16_TO_FP32(x[i*2+0].d));
2534+
const __m128 scale_hi = _mm_set1_ps(GGML_FP16_TO_FP32(x[i*2+1].d));
2535+
__m256 scale = _mm256_set_m128(scale_hi, scale_lo);
2536+
scale = _mm256_mul_ps(scale, _mm256_broadcast_ss(&y[i].d));
25512537

2552-
// Convert int32_t to float
2553-
const __m128 p = _mm_cvtepi32_ps(i32);
2538+
// Load y vector
2539+
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
25542540

2555-
// Apply the scale, and accumulate
2556-
acc = _mm_fmadd_ps(scale, p, acc);
2557-
}
2541+
// Do the product,
2542+
__m256 p = dotMul(bx, by);
2543+
2544+
// Apply the scale, and accumulate
2545+
acc = _mm256_fmadd_ps(scale, p, acc);
25582546
}
25592547

25602548
// Return horizontal sum of the acc vector
2561-
__m128 res = _mm_add_ps(acc, _mm_movehl_ps(acc, acc));
2562-
res = _mm_add_ss(res, _mm_movehdup_ps(res));
2563-
sumf = _mm_cvtss_f32(res);
2549+
sumf = horizontalSum(acc);
25642550
#else
25652551
for (int i = 0; i < nb; i++) {
25662552
const float d0 = GGML_FP16_TO_FP32(x[i].d);
25672553
const float d1 = y[i/2].d;
25682554

2569-
uint64_t qs0 = x[i].qs;
2555+
uint_fast32_t lo0 = x[i].qlo;
2556+
uint_fast32_t hi0 = x[i].qhi << 2;
25702557
const int8_t * restrict p1 = y[i/2].qs + (i%2)*QK3_0;
25712558

25722559
int sumi = 0;
2573-
for (int j = 0; j < QK3_0; j++) {
2574-
const int8_t i0 = (int8_t)(qs0 & 7) - 4;
2575-
const int_fast16_t i1 = p1[j];
2560+
for (int l = 0; l < QK3_0; l++) {
2561+
const int8_t i0 = (int8_t)((lo0 & 3) | ((hi0 & 4) - 4));
2562+
const int_fast16_t i1 = p1[l];
25762563

25772564
sumi += i0 * i1;
25782565

2579-
qs0 >>= 3;
2566+
lo0 >>= 2;
2567+
hi0 >>= 1;
25802568
}
25812569
sumf += d0 * d1 * sumi;
25822570
}
@@ -12064,11 +12052,13 @@ size_t ggml_quantize_q3_0(const float * src, void * dst, int n, int k, int64_t h
1206412052
quantize_row_q3_0(src + j, y, k);
1206512053

1206612054
for (int i = 0; i < nb; i++) {
12067-
uint64_t qs = y[i].qs;
12055+
uint_fast32_t lo = y[i].qlo;
12056+
uint_fast32_t hi = y[i].qhi << 2;
1206812057
for (int l = 0; l < QK3_0; l++) {
12069-
const int8_t vi = qs & 7;
12058+
int8_t vi = (lo & 3) | (hi & 4);
1207012059
hist[vi]++;
12071-
qs >>= 3;
12060+
lo >>= 2;
12061+
hi >>= 1;
1207212062
}
1207312063
}
1207412064
}

0 commit comments

Comments
 (0)