From 982c7cf5cc56a688c628ca1fee98d0cf2ec1b689 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Sun, 11 Jun 2023 12:48:24 +0300 Subject: [PATCH] metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? --- ggml-metal.metal | 123 +++++++++++++++++++++++------------------------ 1 file changed, 59 insertions(+), 64 deletions(-) diff --git a/ggml-metal.metal b/ggml-metal.metal index 617be4a20..23c0803f9 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -634,6 +634,7 @@ typedef struct { half d; // super-block scale for quantized scales half dmin; // super-block scale for quantized mins } block_q2_k; +// 84 bytes / block typedef struct { uint8_t hmask[QK_K/8]; // quants - high bit @@ -641,6 +642,7 @@ typedef struct { uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits half d; // super-block scale } block_q3_k; +// 110 bytes / block typedef struct { half d; // super-block scale for quantized scales @@ -648,6 +650,7 @@ typedef struct { uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits uint8_t qs[QK_K/2]; // 4--bit quants } block_q4_k; +// 144 bytes / block typedef struct { half d; // super-block scale for quantized scales @@ -656,6 +659,7 @@ typedef struct { uint8_t qh[QK_K/8]; // quants, high bit uint8_t qs[QK_K/2]; // quants, low 4 bits } block_q5_k; +// 176 bytes / block typedef struct { uint8_t ql[QK_K/2]; // quants, lower 4 bits @@ -663,6 +667,7 @@ typedef struct { int8_t scales[QK_K/16]; // scales, quantized with 8 bits half d; // super-block scale } block_q6_k; +// 210 bytes / block static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) { uchar4 r; @@ -723,6 +728,7 @@ static void dequantize_row_q3_k(device const block_q3_k * x, device float * y, i const uint32_t kmask2 = 0x0f0f0f0f; uint32_t aux[4]; + thread const int8_t * scales = (thread const int8_t *)aux; for (int i = 0; i < nb; i++) { @@ -739,25 +745,13 @@ static void dequantize_row_q3_k(device const block_q3_k * x, device float * y, i aux[0] = (a[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); aux[1] = (a[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - char4 scales; - - int shift1 = 0; - - int ia = 0; - int is = 4; + int is = 0; float dl; for (int n = 0; n < QK_K; n += 128) { int shift = 0; for (int j = 0; j < 4; ++j) { - const int shift2 = 2*(j/2); - - if (is == 4) { - scales = as_type(aux[ia++]); - is = 0; - } - dl = d_all * (scales[is++] - 32); for (int l = 0; l < 16; ++l) { *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((hm[l+ 0] & m) ? 0 : 4)); @@ -772,7 +766,6 @@ static void dequantize_row_q3_k(device const block_q3_k * x, device float * y, i m <<= 1; } q += 32; - shift1 += 4; } } @@ -1051,29 +1044,17 @@ kernel void kernel_mul_mat_q3_k_f32( device const float * src1, device float * dst, constant int64_t & ne00, - constant int64_t & ne01, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, constant int64_t & ne10, - constant int64_t & ne11, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb12, constant int64_t & ne0, constant int64_t & ne1, threadgroup float * sum [[threadgroup(0)]], uint2 tgpig[[threadgroup_position_in_grid]], - uint2 tpig[[thread_position_in_grid]], // we don't use this for now uint2 tpitg[[thread_position_in_threadgroup]], uint2 tptg[[threads_per_threadgroup]]) { const uint32_t kmask1 = 0x03030303; const uint32_t kmask2 = 0x0f0f0f0f; - uint32_t aux[2]; - - const uint8_t m1 = 1; const uint8_t m3 = 3; const int8_t m4 = 4; @@ -1088,61 +1069,75 @@ kernel void kernel_mul_mat_q3_k_f32( const int nth = tptg.x*tptg.y; const int ith = tptg.y*tpitg.x + tpitg.y; - const int step = QK_K / tptg.y; // we expect this to be 16 - const int iqs = step * tpitg.y; // 0...240 in steps of 16 - const int ip = iqs / 128; // 0 or 1 - const int il = (iqs - 128*ip)/16; // 0...7 + const int tid = tpitg.y; + const int il = tid/4; // 0...3 0 -> 0...63, 1 -> 64...127, 2 -> 128...191, 3 -> 192...255 + const int ip = il / 2; // 0 or 1 0 -> use 1st 32 q's (0...127), 1 -> 2nd 32 (128...255) + const int is = il % 2; // 0 or 1 0 -> 0...63, 128...191, 1 -> 64...127, 192...255 + const int ir = tid - 4*il; // 0...3 const int n = 4; - const int l0 = n * il; - const int is = l0/16; - const uint8_t m = m1 << (4*ip); - const uchar4 mask = {m, (uint8_t)(m << 1), (uint8_t)(m << 2), (uint8_t)(m << 3)}; + const int l0 = n * ir; // first index for this thread within a group of 32 (0, 4, 8, 12) + // 0...31 use 1<<0, 32...63 use 1<<1, 64...95 use 1<<2, 96...128 use 1<<3, etc. + // we process 64*il...64*il+63 -> 1st mask is 1<<(2*il), second is 1<<(2*il+1) + // masks for high bit + const uint8_t m = 1 << (2*il); + const uchar2 mask = {m, (uint8_t)(m << 1)}; - const int shift1 = 4*ip; - const int shift2 = shift1 + 2; + const int shift1 = 4*ip; // 1st shift for scale. must be 0 (0...127) or 4 (128...255) + const int shift2 = 2*il; // 2nd shift for scale. 0, 2, 4, or 6 + // 1st shift for quants must be 0 in 0...31, 2 in 32...64, 4 in 64...96, 6 in 96...128, then agsin 0, 2, etc. + const int shift3 = 4*is; + const int shift4 = shift3 + 2; - //int8_t sc[4]; + const int q_offset = 32*ip + l0; + const int y_offset = 64*il + l0; float sumf = 0; for (int i = tpitg.x; i < nb; i += tptg.x) { + // Copied from the C de-quantization code + //aux[0] = ((a[0] >> 0) & kmask2) | (((a[2] >> 0) & kmask1) << 4); + //aux[1] = ((a[1] >> 0) & kmask2) | (((a[2] >> 2) & kmask1) << 4); + //aux[2] = ((a[0] >> 4) & kmask2) | (((a[2] >> 4) & kmask1) << 4); + //aux[3] = ((a[1] >> 4) & kmask2) | (((a[2] >> 6) & kmask1) << 4); + + //// 0....63 we need a[0] with shift=0, a[2] with shift 0 + //// 64...127 we need a[1] with shift=0, a[2] with shift 2 + ////128...191 we need a[0] with shift=4, a[2] with shift 4 + ////192...255 we need a[1] with shift=4, a[2] with shift 6 + //// a[is] >> (4*ip) & 0xF | a[2] >> (2*il) & 3 device const uint32_t * a = (device const uint32_t *)x[i].scales; - uint32_t tmp = a[2]; - aux[0] = ((a[0] >> shift1) & kmask2) | (((tmp >> shift1) & kmask1) << 4); - aux[1] = ((a[1] >> shift1) & kmask2) | (((tmp >> shift2) & kmask1) << 4); + const char4 sc = as_type(((a[is] >> shift1) & kmask2) | (((a[2] >> shift2) & kmask1) << 4)); - device const uint8_t * q = x[i].qs + 32*ip + l0; - device const uint8_t * hm = x[i].hmask + l0; - //device const uint8_t * scales = x[i].scales; + // Here I was thinking "what if the above is not processed correctly because x[i].scales is not 4-byte + // aligned?". If that was the issue, using a uint16_t pointer should solve it as x[i].scales is 2-byte aligned. + // It does not solve the problem, it just makes it run slower. + //device const uint16_t * a = (device const uint16_t *)x[i].scales; + //const char2 sc1 = as_type((uint16_t)(((a[2*is+0] >> shift1) & kmask2) | (((a[4] >> shift2) & kmask1) << 4))); + //const char2 sc2 = as_type((uint16_t)(((a[2*is+1] >> shift1) & kmask2) | (((a[5] >> shift2) & kmask1) << 4))); - device const float * y = yy + i * QK_K + 128*ip + l0; + device const uint8_t * q = x[i].qs + q_offset; + device const uint8_t * h = x[i].hmask + l0; - const float dall = x[i].d; + device const float * y = yy + i * QK_K + y_offset; - const char4 sc1 = as_type(aux[0]); - const char4 sc2 = as_type(aux[1]); - - //sc[0] = ((scales[is+0] >> shift1) & 0xF) | (((scales[is+ 8] >> shift1) & m3) << 4); - //sc[1] = ((scales[is+2] >> shift1) & 0xF) | (((scales[is+10] >> shift1) & m3) << 4); - //sc[2] = ((scales[is+4] >> shift1) & 0xF) | (((scales[is+ 8] >> shift2) & m3) << 4); - //sc[3] = ((scales[is+6] >> shift1) & 0xF) | (((scales[is+10] >> shift2) & m3) << 4); + const float dall = (float)x[i].d; float4 sums = {0.f, 0.f, 0.f, 0.f}; for (int l = 0; l < n; ++l) { - sums[0] += y[l+ 0] * ((int8_t)((q[l] >> 0) & m3) - ((hm[l] & mask[0]) ? 0 : m4)); - sums[1] += y[l+32] * ((int8_t)((q[l] >> 2) & m3) - ((hm[l] & mask[1]) ? 0 : m4)); - sums[2] += y[l+64] * ((int8_t)((q[l] >> 4) & m3) - ((hm[l] & mask[2]) ? 0 : m4)); - sums[3] += y[l+96] * ((int8_t)((q[l] >> 6) & m3) - ((hm[l] & mask[3]) ? 0 : m4)); + sums[0] += y[l+ 0] * ((int8_t)((q[l+ 0] >> shift3) & m3) - ((h[l+ 0] & mask[0]) ? 0 : m4)); + sums[1] += y[l+16] * ((int8_t)((q[l+16] >> shift3) & m3) - ((h[l+16] & mask[0]) ? 0 : m4)); + sums[2] += y[l+32] * ((int8_t)((q[l+ 0] >> shift4) & m3) - ((h[l+ 0] & mask[1]) ? 0 : m4)); + sums[3] += y[l+48] * ((int8_t)((q[l+16] >> shift4) & m3) - ((h[l+16] & mask[1]) ? 0 : m4)); } - //sumf += dall * (sums[0] * (sc[0] - 32) - // + sums[1] * (sc[1] - 32) - // + sums[2] * (sc[2] - 32) - // + sums[3] * (sc[3] - 32)); - sumf += dall * (sums[0] * (sc1[is+0] - 32) - + sums[1] * (sc1[is+2] - 32) - + sums[2] * (sc2[is+0] - 32) - + sums[3] * (sc2[is+2] - 32)); + sumf += dall * (sums[0] * (sc[0] - 32) + + sums[1] * (sc[1] - 32) + + sums[2] * (sc[2] - 32) + + sums[3] * (sc[3] - 32)); + //sumf += dall * (sums[0] * (sc1[0] - 32) + // + sums[1] * (sc1[1] - 32) + // + sums[2] * (sc2[0] - 32) + // + sums[3] * (sc2[1] - 32)); }