ngxson/FAQ.md

Last active August 5, 2025 17:29

Star (129) You must be signed in to star a gist
Fork (22) You must be signed in to fork a gist

Learn more about clone URLs
Clone this repository at <script src="https://gist.github.com/ngxson/307140d24d80748bd683b396ba13be07.js"></script>
Save ngxson/307140d24d80748bd683b396ba13be07 to your computer and use it in GitHub Desktop.

Download ZIP

convert ARM NEON to WASM SIMD prompt

Raw

FAQ.md

Why did you do this?

Relax, I only have one Sunday to work on idea, literally my weekend project. So I tried Deepseek to see if it can help. Surprisingly, it works and it saves me another weekend...

What is your setup?

Just chat.deepseek.com (cost = free) with prompts adapted from this gist.

Does it work in one-shot or I have to prompt it multiple times?

For the qX_0 variants, they are actually quite straight-forward so deepseek can come up with a correct result in 1 shot.
For the qX_K it's more complicated, I would say most of the time I need to re-prompt it 4 to 8 more times.
The most difficult was q6_K, the code never works until I ask it to only optimize one specific part, while leaving the rest intact (so it does not mess up everything)

It only does conversion ARM NEON --> WASM SIMD, or it can invent new WASM SIMD code from scratch?

It can do both. For qX_0 I asked it to convert, and for qX_K I asked it to invent new code.

How much time does it spent to think?

Around 3-5 minutes per response.

Prompt is very long, what to do?

You can condense the prompt to this format, then create a new conversation (see example at the bottom):

Problem description
"Here is your last failed attempt, improve from this: [paste the last generated code here]"

Raw

prompt.md

Your task is to convert a given C++ ARM NEON SIMD to WASM SIMD. Here is an example of another function:

void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;

    int ib = 0;
    float sumf = 0;

    assert(n % qk == 0);
    assert(qk == QK5_0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q5_0 * restrict x = vx;
    const block_q8_0 * restrict y = vy;

#if defined(__ARM_NEON)
    float32x4_t sumv0 = vdupq_n_f32(0.0f);
    float32x4_t sumv1 = vdupq_n_f32(0.0f);

    uint32_t qh0;
    uint32_t qh1;

    uint64_t tmp0[4];
    uint64_t tmp1[4];

    for (; ib + 1 < nb; ib += 2) {
        const block_q5_0 * restrict x0 = &x[ib];
        const block_q5_0 * restrict x1 = &x[ib + 1];
        const block_q8_0 * restrict y0 = &y[ib];
        const block_q8_0 * restrict y1 = &y[ib + 1];

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

        // extract the 5th bit via lookup table ((!b) << 4)
        memcpy(&qh0, x0->qh, sizeof(qh0));
        memcpy(&qh1, x1->qh, sizeof(qh1));

        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];

        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];

        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));

        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
        const uint8x16_t v0_1 = vld1q_u8(x1->qs);

        // 4-bit -> 8-bit
        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));

        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);

        // load y
        const int8x16_t v1_0l = vld1q_s8(y0->qs);
        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
        const int8x16_t v1_1l = vld1q_s8(y1->qs);
        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__wasm_simd128__)
    v128_t sumv = wasm_f32x4_splat(0.0f);

    uint32_t qh;
    uint64_t tmp[4];

    // TODO: check if unrolling this is better
    for (; ib < nb; ++ib) {
        const block_q5_0 * restrict x0 = &x[ib];
        const block_q8_0 * restrict y0 = &y[ib];

        const v128_t m4b  = wasm_i8x16_splat(0x0F);

        // extract the 5th bit
        memcpy(&qh, x0->qh, sizeof(qh));

        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
        tmp[3] = table_b2b_1[(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);

        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
        const v128_t v0hf = wasm_i8x16_sub(v0h, 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);

        // 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(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
    }

    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#endif

    for (; ib < nb; ++ib) {
        uint32_t qh;
        memcpy(&qh, x[ib].qh, sizeof(qh));

        int sumi0 = 0;
        int sumi1 = 0;

        for (int j = 0; j < qk/2; ++j) {
            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));

            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);

            sumi0 += (x0 * y[ib].qs[j]);
            sumi1 += (x1 * y[ib].qs[j + qk/2]);
        }

        int sumi = sumi0 + sumi1;
        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
}

Here is the function that you need to convert:

void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;

    assert(n % qk == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q8_0 * restrict x = vx;
    const block_q8_0 * restrict y = vy;

    int ib = 0;
    float sumf = 0;

#if defined(__ARM_NEON)
    float32x4_t sumv0 = vdupq_n_f32(0.0f);
    float32x4_t sumv1 = vdupq_n_f32(0.0f);

    for (; ib + 1 < nb; ib += 2) {
        const block_q8_0 * restrict x0 = &x[ib + 0];
        const block_q8_0 * restrict x1 = &x[ib + 1];
        const block_q8_0 * restrict y0 = &y[ib + 0];
        const block_q8_0 * restrict y1 = &y[ib + 1];

        const int8x16_t x0_0 = vld1q_s8(x0->qs);
        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
        const int8x16_t x1_0 = vld1q_s8(x1->qs);
        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);

        // load y
        const int8x16_t y0_0 = vld1q_s8(y0->qs);
        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
        const int8x16_t y1_0 = vld1q_s8(y1->qs);
        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));

        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#endif
    for (; ib < nb; ++ib) {
        int sumi = 0;

        for (int j = 0; j < qk; j++) {
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
}

You must start your code with #elif defined(__wasm_simd128__)

Raw

prompt2.md

Your task is to convert a given C code SIMD to WASM SIMD. Here is an example of another function:

void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q4_K * restrict x = vx;
    const block_q8_K * restrict y = vy;

    const int nb = n / QK_K;

    static const uint32_t kmask1 = 0x3f3f3f3f;
    static const uint32_t kmask2 = 0x0f0f0f0f;
    static const uint32_t kmask3 = 0x03030303;

    uint32_t utmp[4];

#ifdef __ARM_NEON
    const uint8x16_t m4b = vdupq_n_u8(0xf);
    const int32x4_t mzero = vdupq_n_s32(0);

    ggml_int8x16x2_t q4bytes;
    ggml_int8x16x2_t q8bytes;

    float sumf = 0;

    for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);

        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));

        memcpy(utmp, x[i].scales, 12);

        uint32x2_t mins8 = { 0 };
        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);

        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[0] &= kmask1;

        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
        sumf -= dmin * vaddvq_s32(prod);

        const uint8_t * scales = (const uint8_t *)utmp;

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        int32_t sumi1 = 0;
        int32_t sumi2 = 0;

        for (int j = 0; j < QK_K/64; ++j) {
            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;

            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));

            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
            sumi1 += vaddvq_s32(p1) * scales[2*j+0];

            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));

            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);

            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
        }

        sumf += d * (sumi1 + sumi2);

    }

    *s = sumf;

#elif defined(__wasm_simd128__)
    // WASM SIMD128 implementation
    const uint8_t * scales = (const uint8_t*)&utmp[0];
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        // Process scales and mins
        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        // Sum mins * q8sums
        int32_t sumi = 0;
        const int16_t * restrict q8sums = y[i].bsums;
        const uint8_t * m = (const uint8_t *)&utmp[2];
        for (int j = 0; j < 16; j += 2) {
            sumi += (q8sums[j] + q8sums[j+1]) * m[j/2];
        }
        sumf -= dmin * sumi;

        int32_t sumi1 = 0;
        int32_t sumi2 = 0;

        for (int j = 0; j < QK_K/64; ++j) {
            // Load 64 4-bit weights (32 bytes)
            const v128_t q4x0 = wasm_v128_load(q4);
            const v128_t q4x1 = wasm_v128_load(q4 + 16);
            q4 += 32;

            // Split into low/high nibbles
            const v128_t q4l0 = wasm_v128_and(q4x0, wasm_i8x16_splat(0x0F));
            const v128_t q4h0 = wasm_u8x16_shr(q4x0, 4);
            const v128_t q4l1 = wasm_v128_and(q4x1, wasm_i8x16_splat(0x0F));
            const v128_t q4h1 = wasm_u8x16_shr(q4x1, 4);

            // Load 64 8-bit values (64 bytes)
            const v128_t q8x0 = wasm_v128_load(q8);
            const v128_t q8x1 = wasm_v128_load(q8 + 16);
            const v128_t q8x2 = wasm_v128_load(q8 + 32);
            const v128_t q8x3 = wasm_v128_load(q8 + 48);
            q8 += 64;

            // Low nibble products
            v128_t vacc1 = wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4l0),
                wasm_i16x8_extend_low_i8x16(q8x0)
            );
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4l0),
                wasm_i16x8_extend_high_i8x16(q8x0)
            ));
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4l1),
                wasm_i16x8_extend_low_i8x16(q8x1)
            ));
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4l1),
                wasm_i16x8_extend_high_i8x16(q8x1)
            ));

            // High nibble products
            v128_t vacc2 = wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4h0),
                wasm_i16x8_extend_low_i8x16(q8x2)
            );
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4h0),
                wasm_i16x8_extend_high_i8x16(q8x2)
            ));
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4h1),
                wasm_i16x8_extend_low_i8x16(q8x3)
            ));
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4h1),
                wasm_i16x8_extend_high_i8x16(q8x3)
            ));

            // Accumulate scaled results
            int32_t vacc1_sum = wasm_i32x4_extract_lane(vacc1, 0) + wasm_i32x4_extract_lane(vacc1, 1) +
                                wasm_i32x4_extract_lane(vacc1, 2) + wasm_i32x4_extract_lane(vacc1, 3);
            sumi1 += vacc1_sum * scales[2*j];

            int32_t vacc2_sum = wasm_i32x4_extract_lane(vacc2, 0) + wasm_i32x4_extract_lane(vacc2, 1) +
                                wasm_i32x4_extract_lane(vacc2, 2) + wasm_i32x4_extract_lane(vacc2, 3);
            sumi2 += vacc2_sum * scales[2*j+1];
        }

        sumf += d * (sumi1 + sumi2);
    }

    *s = sumf;

#elif defined __AVX__

    const __m128i m4 = _mm_set1_epi8(0xF);
    const __m128i m2 = _mm_set1_epi8(0x2);

    __m256 acc = _mm256_setzero_ps();
    __m128 acc_m = _mm_setzero_ps();

   for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
        const __m128i scales = _mm_cvtepu8_epi16(utmps);
        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));

        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
        const __m128i prod = _mm_madd_epi16(mins, q8s);
        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);

        __m128i sumi_0 = _mm_setzero_si128();
        __m128i sumi_1 = _mm_setzero_si128();

        __m128i shuffle = _mm_set1_epi16(0x0100);
        for (int j = 0; j < QK_K/64; ++j) {

            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
            shuffle = _mm_add_epi16(shuffle, m2);
            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
            shuffle = _mm_add_epi16(shuffle, m2);

            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);

            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
            p16l = _mm_madd_epi16(scale_l, p16l);
            sumi_0 = _mm_add_epi32(sumi_0, p16l);
            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
            p16l = _mm_madd_epi16(scale_l, p16l);
            sumi_1 = _mm_add_epi32(sumi_1, p16l);

            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
            p16h = _mm_madd_epi16(scale_h, p16h);
            sumi_0 = _mm_add_epi32(sumi_0, p16h);
            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
            p16h = _mm_madd_epi16(scale_h, p16h);
            sumi_1 = _mm_add_epi32(sumi_1, p16h);

        }

        __m256 vd = _mm256_set1_ps(d);
        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);

    }

    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));

    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);

#else

    const uint8_t * scales = (const uint8_t*)&utmp[0];
    const uint8_t * mins   = (const uint8_t*)&utmp[2];

    int8_t  aux8[QK_K];
    int16_t aux16[8];
    float   sums [8];
    int32_t aux32[8];
    memset(sums, 0, 8*sizeof(float));

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
        const uint8_t * restrict q4 = x[i].qs;
        const  int8_t * restrict q8 = y[i].qs;
        memset(aux32, 0, 8*sizeof(int32_t));
        int8_t * restrict a = aux8;
        for (int j = 0; j < QK_K/64; ++j) {
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
            a += 32;
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
            a += 32; q4 += 32;
        }
        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        int sumi = 0;
        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
        a = aux8;
        int is = 0;
        for (int j = 0; j < QK_K/32; ++j) {
            int32_t scale = scales[is++];
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
#endif
}

Here is a function. You need to convert it to WASM SIMD.

void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q6_K * restrict x = vx;
    const block_q8_K * restrict y = vy;

    const int nb = n / QK_K;

#ifdef __ARM_NEON
    float sum = 0;

    const uint8x16_t m4b = vdupq_n_u8(0xF);
    const int32x4_t  vzero = vdupq_n_s32(0);
    //const int8x16_t  m32s = vdupq_n_s8(32);

    const uint8x16_t mone = vdupq_n_u8(3);

    ggml_int8x16x4_t q6bytes;
    ggml_uint8x16x4_t q6h;

    for (int i = 0; i < nb; ++i) {

        const float d_all = GGML_FP16_TO_FP32(x[i].d);

        const uint8_t * restrict q6 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const int8_t  * restrict q8 = y[i].qs;

        const int8_t * restrict scale = x[i].scales;

        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
        const int8x16_t scales = vld1q_s8(scale);
        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};

        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
        int32_t isum_mins = vaddvq_s32(prod);

        int32_t isum = 0;

        for (int j = 0; j < QK_K/128; ++j) {

            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;

            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 2);
            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);

            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));

            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];

            scale += 4;

            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;

            shifted = vshrq_n_u8(qhbits.val[0], 4);
            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 4);
            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[0], 6);
            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 6);
            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);

            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));

            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
            scale += 4;
        }
        //sum += isum * d_all * y[i].d;
        sum += d_all * y[i].d * (isum - 32 * isum_mins);

    }
    *s = sum;

#elif defined __AVX__

    const __m128i m3 = _mm_set1_epi8(3);
    const __m128i m15 = _mm_set1_epi8(15);

    __m256 acc = _mm256_setzero_ps();

    for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);

        const uint8_t * restrict q4 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const int8_t  * restrict q8 = y[i].qs;

        // handle the q6_k -32 offset separately using bsums
        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);

        __m128i sumi_0 = _mm_setzero_si128();
        __m128i sumi_1 = _mm_setzero_si128();

        int is = 0;

        for (int j = 0; j < QK_K/128; ++j) {

            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;

            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);

            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;

            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);

            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;

            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);

            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
            is += 4;

            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);

            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));

        }

        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
    }

    *s = hsum_float_8(acc);

#else

    int8_t  aux8[QK_K];
    int16_t aux16[8];
    float   sums [8];
    int32_t aux32[8];
    memset(sums, 0, 8*sizeof(float));

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
        const uint8_t * restrict q4 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const  int8_t * restrict q8 = y[i].qs;
        memset(aux32, 0, 8*sizeof(int32_t));
        int8_t * restrict a = aux8;
        for (int j = 0; j < QK_K; j += 128) {
            for (int l = 0; l < 32; ++l) {
                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
            }
            a  += 128;
            q4 += 64;
            qh += 32;
        }
        a = aux8;
        int is = 0;
        for (int j = 0; j < QK_K/16; ++j) {
            int scale = x[i].scales[is++];
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
#endif
}

You must start your code with #elif defined(__wasm_simd128__)

To think about it, you need to take into account both the refenrence code from ARM NEON and AVX implementation.

Raw

prompt3.md

Here is a function in C, you have to convert it to WASM SIMD 128:

You must start your code with #elif defined(__wasm_simd128__)

To think about it, you need to take into account both the refenrence code from ARM NEON and AVX implementation.

Please note that this is YOUR last attempt. It compiles, but give inaccurate result. Improve from that.

You must start your code with #elif defined(__wasm_simd128__)

To think about it, you need to take into account both the refenrence code from ARM NEON, riscv_v_intrinsic and other implementation that you can see on the code. Make sure the output value is as accurate as possible.

Raw

prompt_from_scratch.md

Here is a function in C, you have to convert it to WASM SIMD 128.

This function is prone to produce inaccuracy result, it is very sensitive. Only optimize part that you are absolutely sure.

Think carefully about it. Hint, focus more on the loop with aux16, aux32

void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q6_K * restrict x = vx;
    const block_q8_K * restrict y = vy;

    const int nb = n / QK_K;

    int8_t  aux8[QK_K];
    int16_t aux16[8];
    float   sums [8];
    int32_t aux32[8];
    memset(sums, 0, 8*sizeof(float));

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
        const uint8_t * restrict q4 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const  int8_t * restrict q8 = y[i].qs;
        memset(aux32, 0, 8*sizeof(int32_t));
        int8_t * restrict a = aux8;
        for (int j = 0; j < QK_K; j += 128) {
            for (int l = 0; l < 32; ++l) {
                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
            }
            a  += 128;
            q4 += 64;
            qh += 32;
        }
        a = aux8;
        int is = 0;
        for (int j = 0; j < QK_K/16; ++j) {
            int scale = x[i].scales[is++];
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
}

Raw

prompt_hard.md

NOTE: This task (q6_K_q8_K) is HARD. The prompt below always produces failed result

Your task is to convert a given C code SIMD to WASM SIMD. Here is an example of another function:

void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q4_K * restrict x = vx;
    const block_q8_K * restrict y = vy;

    const int nb = n / QK_K;

    static const uint32_t kmask1 = 0x3f3f3f3f;
    static const uint32_t kmask2 = 0x0f0f0f0f;
    static const uint32_t kmask3 = 0x03030303;

    uint32_t utmp[4];

#ifdef __ARM_NEON
    const uint8x16_t m4b = vdupq_n_u8(0xf);
    const int32x4_t mzero = vdupq_n_s32(0);

    ggml_int8x16x2_t q4bytes;
    ggml_int8x16x2_t q8bytes;

    float sumf = 0;

    for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);

        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));

        memcpy(utmp, x[i].scales, 12);

        uint32x2_t mins8 = { 0 };
        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);

        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[0] &= kmask1;

        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
        sumf -= dmin * vaddvq_s32(prod);

        const uint8_t * scales = (const uint8_t *)utmp;

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        int32_t sumi1 = 0;
        int32_t sumi2 = 0;

        for (int j = 0; j < QK_K/64; ++j) {
            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;

            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));

            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
            sumi1 += vaddvq_s32(p1) * scales[2*j+0];

            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));

            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);

            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
        }

        sumf += d * (sumi1 + sumi2);

    }

    *s = sumf;

#elif defined(__wasm_simd128__)
    // WASM SIMD128 implementation
    const uint8_t * scales = (const uint8_t*)&utmp[0];
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        // Process scales and mins
        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        // Sum mins * q8sums
        int32_t sumi = 0;
        const int16_t * restrict q8sums = y[i].bsums;
        const uint8_t * m = (const uint8_t *)&utmp[2];
        for (int j = 0; j < 16; j += 2) {
            sumi += (q8sums[j] + q8sums[j+1]) * m[j/2];
        }
        sumf -= dmin * sumi;

        int32_t sumi1 = 0;
        int32_t sumi2 = 0;

        for (int j = 0; j < QK_K/64; ++j) {
            // Load 64 4-bit weights (32 bytes)
            const v128_t q4x0 = wasm_v128_load(q4);
            const v128_t q4x1 = wasm_v128_load(q4 + 16);
            q4 += 32;

            // Split into low/high nibbles
            const v128_t q4l0 = wasm_v128_and(q4x0, wasm_i8x16_splat(0x0F));
            const v128_t q4h0 = wasm_u8x16_shr(q4x0, 4);
            const v128_t q4l1 = wasm_v128_and(q4x1, wasm_i8x16_splat(0x0F));
            const v128_t q4h1 = wasm_u8x16_shr(q4x1, 4);

            // Load 64 8-bit values (64 bytes)
            const v128_t q8x0 = wasm_v128_load(q8);
            const v128_t q8x1 = wasm_v128_load(q8 + 16);
            const v128_t q8x2 = wasm_v128_load(q8 + 32);
            const v128_t q8x3 = wasm_v128_load(q8 + 48);
            q8 += 64;

            // Low nibble products
            v128_t vacc1 = wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4l0),
                wasm_i16x8_extend_low_i8x16(q8x0)
            );
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4l0),
                wasm_i16x8_extend_high_i8x16(q8x0)
            ));
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4l1),
                wasm_i16x8_extend_low_i8x16(q8x1)
            ));
            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4l1),
                wasm_i16x8_extend_high_i8x16(q8x1)
            ));

            // High nibble products
            v128_t vacc2 = wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4h0),
                wasm_i16x8_extend_low_i8x16(q8x2)
            );
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4h0),
                wasm_i16x8_extend_high_i8x16(q8x2)
            ));
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_low_i8x16(q4h1),
                wasm_i16x8_extend_low_i8x16(q8x3)
            ));
            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
                wasm_i16x8_extend_high_i8x16(q4h1),
                wasm_i16x8_extend_high_i8x16(q8x3)
            ));

            // Accumulate scaled results
            int32_t vacc1_sum = wasm_i32x4_extract_lane(vacc1, 0) + wasm_i32x4_extract_lane(vacc1, 1) +
                                wasm_i32x4_extract_lane(vacc1, 2) + wasm_i32x4_extract_lane(vacc1, 3);
            sumi1 += vacc1_sum * scales[2*j];

            int32_t vacc2_sum = wasm_i32x4_extract_lane(vacc2, 0) + wasm_i32x4_extract_lane(vacc2, 1) +
                                wasm_i32x4_extract_lane(vacc2, 2) + wasm_i32x4_extract_lane(vacc2, 3);
            sumi2 += vacc2_sum * scales[2*j+1];
        }

        sumf += d * (sumi1 + sumi2);
    }

    *s = sumf;

#elif defined __AVX__

    const __m128i m4 = _mm_set1_epi8(0xF);
    const __m128i m2 = _mm_set1_epi8(0x2);

    __m256 acc = _mm256_setzero_ps();
    __m128 acc_m = _mm_setzero_ps();

   for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);

        const uint8_t * restrict q4 = x[i].qs;
        const int8_t  * restrict q8 = y[i].qs;

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
        const __m128i scales = _mm_cvtepu8_epi16(utmps);
        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));

        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
        const __m128i prod = _mm_madd_epi16(mins, q8s);
        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);

        __m128i sumi_0 = _mm_setzero_si128();
        __m128i sumi_1 = _mm_setzero_si128();

        __m128i shuffle = _mm_set1_epi16(0x0100);
        for (int j = 0; j < QK_K/64; ++j) {

            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
            shuffle = _mm_add_epi16(shuffle, m2);
            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
            shuffle = _mm_add_epi16(shuffle, m2);

            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);

            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
            p16l = _mm_madd_epi16(scale_l, p16l);
            sumi_0 = _mm_add_epi32(sumi_0, p16l);
            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
            p16l = _mm_madd_epi16(scale_l, p16l);
            sumi_1 = _mm_add_epi32(sumi_1, p16l);

            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
            p16h = _mm_madd_epi16(scale_h, p16h);
            sumi_0 = _mm_add_epi32(sumi_0, p16h);
            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
            p16h = _mm_madd_epi16(scale_h, p16h);
            sumi_1 = _mm_add_epi32(sumi_1, p16h);

        }

        __m256 vd = _mm256_set1_ps(d);
        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);

    }

    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));

    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);

#else

    const uint8_t * scales = (const uint8_t*)&utmp[0];
    const uint8_t * mins   = (const uint8_t*)&utmp[2];

    int8_t  aux8[QK_K];
    int16_t aux16[8];
    float   sums [8];
    int32_t aux32[8];
    memset(sums, 0, 8*sizeof(float));

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
        const uint8_t * restrict q4 = x[i].qs;
        const  int8_t * restrict q8 = y[i].qs;
        memset(aux32, 0, 8*sizeof(int32_t));
        int8_t * restrict a = aux8;
        for (int j = 0; j < QK_K/64; ++j) {
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
            a += 32;
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
            a += 32; q4 += 32;
        }
        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;

        int sumi = 0;
        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
        a = aux8;
        int is = 0;
        for (int j = 0; j < QK_K/32; ++j) {
            int32_t scale = scales[is++];
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
#endif
}

Here is a function. You need to convert it to WASM SIMD.

void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q6_K * restrict x = vx;
    const block_q8_K * restrict y = vy;

    const int nb = n / QK_K;

#ifdef __ARM_NEON
    float sum = 0;

    const uint8x16_t m4b = vdupq_n_u8(0xF);
    const int32x4_t  vzero = vdupq_n_s32(0);
    //const int8x16_t  m32s = vdupq_n_s8(32);

    const uint8x16_t mone = vdupq_n_u8(3);

    ggml_int8x16x4_t q6bytes;
    ggml_uint8x16x4_t q6h;

    for (int i = 0; i < nb; ++i) {

        const float d_all = GGML_FP16_TO_FP32(x[i].d);

        const uint8_t * restrict q6 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const int8_t  * restrict q8 = y[i].qs;

        const int8_t * restrict scale = x[i].scales;

        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
        const int8x16_t scales = vld1q_s8(scale);
        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};

        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
        int32_t isum_mins = vaddvq_s32(prod);

        int32_t isum = 0;

        for (int j = 0; j < QK_K/128; ++j) {

            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;

            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 2);
            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);

            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));

            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];

            scale += 4;

            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;

            shifted = vshrq_n_u8(qhbits.val[0], 4);
            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 4);
            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[0], 6);
            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
            shifted = vshrq_n_u8(qhbits.val[1], 6);
            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);

            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));

            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
            scale += 4;
        }
        //sum += isum * d_all * y[i].d;
        sum += d_all * y[i].d * (isum - 32 * isum_mins);

    }
    *s = sum;

#elif defined __AVX__

    const __m128i m3 = _mm_set1_epi8(3);
    const __m128i m15 = _mm_set1_epi8(15);

    __m256 acc = _mm256_setzero_ps();

    for (int i = 0; i < nb; ++i) {

        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);

        const uint8_t * restrict q4 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const int8_t  * restrict q8 = y[i].qs;

        // handle the q6_k -32 offset separately using bsums
        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);

        __m128i sumi_0 = _mm_setzero_si128();
        __m128i sumi_1 = _mm_setzero_si128();

        int is = 0;

        for (int j = 0; j < QK_K/128; ++j) {

            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;

            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);

            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;

            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);

            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;

            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);

            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
            is += 4;

            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);

            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));

        }

        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
    }

    *s = hsum_float_8(acc);

#else

    int8_t  aux8[QK_K];
    int16_t aux16[8];
    float   sums [8];
    int32_t aux32[8];
    memset(sums, 0, 8*sizeof(float));

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
        const uint8_t * restrict q4 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
        const  int8_t * restrict q8 = y[i].qs;
        memset(aux32, 0, 8*sizeof(int32_t));
        int8_t * restrict a = aux8;
        for (int j = 0; j < QK_K; j += 128) {
            for (int l = 0; l < 32; ++l) {
                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
            }
            a  += 128;
            q4 += 64;
            qh += 32;
        }
        a = aux8;
        int is = 0;
        for (int j = 0; j < QK_K/16; ++j) {
            int scale = x[i].scales[is++];
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
#endif
}

You must start your code with #elif defined(__wasm_simd128__)

To think about it, you need to take into account both the refenrence code from ARM NEON and AVX implementation.

Author

ngxson commented Jan 27, 2025

For the qX_0 variants, they are actually quite straight-forward so deepseek can come up with a correct result in 1 shot

For the qX_K it's more complicated, I would say most of the time I need to re-prompt it 4 to 8 more times.

The most difficult was q6_K, the code never works until I ask it to only optimize one part, but not all (so it does not mess up everything)

henryclw commented Jan 27, 2025

Thank you so much for sharing this

timwu-ipevo commented Jan 27, 2025

Could you share the thought process R1 outputted? I'd like to understand how it approached this complex problem with just a single example.

Author

ngxson commented Jan 27, 2025 •

edited

Loading

@timwu-ipevo It's too long so I don't think it's worth sharing here. You can try it yourself on chat.deepseek.com, it's free

Manamama commented Jan 28, 2025

Amazing results, ta too. FYI, I have landed here from a Reddit llama thread discussing some Deepseek's geopolitics...

0xCA commented Jan 28, 2025 •

edited

Loading

Which qX_0 exactly you used? I can only see q8_0 here or here
You also said q6_K was the worst. Does it mean anything lower than 6bit is even worse?

Author

ngxson commented Jan 28, 2025

@0xCA All the qX_0_q8_0 functions were generated by that one prompt, just replace the data.

q6_K is an exception case because I need to ask it to reinvent the whole thing from scratch. Not sure why, but other qX_K requires less trials to get it right. For example, q4_K just work one-shot IIRC

Author

ngxson commented Jan 28, 2025

q6_K is an exception case because I need to ask it to reinvent the whole thing from scratch

And btw, for this exact reason, you will see that the code for q6_K looks different from the rest. For example, that's the only place it uses __attribute__((aligned(16)))

0xCA commented Jan 28, 2025

@ngxson ah, seems like I misunderstood you! You meant functions, and I thought you used quantized models of R1.

Author

ngxson commented Jan 28, 2025

No, as stated in the FAQ, I used chat.deepseek.com

And ironically, I came up with this whole idea while optimizing wllama to run deepseek-r1-distilled-qwen-1.5B faster. So the bigger deepseek helping optimize code to run the smaller deepseek.

hjortholm commented Jan 29, 2025

The code is out of my league, but I might have an improvement for the prompt. Try this:

Goal: Translate the ggml_vec_dot_q8_0_q8_0 function from ARM NEON to WASM SIMD, matching the style shown in the provided ggml_vec_dot_q5_0_q8_0 conversion.

Instructions:

Provide a WASM SIMD version of the function.
Begin the code block with #elif defined(wasm_simd128).
Preserve key operations:
Loading data into 128-bit registers (v128_t in WASM).
Performing dot products using WASM intrinsics analogous to vmlaq_n_f32 and vcvtq_f32_s32.
Summation in float32 format, as in the example.
Ensure the final WASM SIMD code is self-contained and references the same arrays/structs (block_q8_0, qs, d) as the ARM version.
Clearly comment any lines where NEON intrinsics have a direct WASM SIMD equivalent.
Retain the function signature and the final accumulation (sumf) approach.
If loops are unrolled similarly to the example, replicate that strategy. If performance trade-offs exist, note them in the comments.
Output Criteria:

A complete WASM SIMD function ready to place in the code as the counterpart to the NEON block.
Comments highlighting any differences or special considerations for WebAssembly.
Maintenance of the same input parameters and usage patterns, ensuring that the rest of the code remains compatible.

Here is an example of another function:

void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;

    int ib = 0;
    float sumf = 0;

    assert(n % qk == 0);
    assert(qk == QK5_0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q5_0 * restrict x = vx;
    const block_q8_0 * restrict y = vy;

#if defined(__ARM_NEON)
    float32x4_t sumv0 = vdupq_n_f32(0.0f);
    float32x4_t sumv1 = vdupq_n_f32(0.0f);

    uint32_t qh0;
    uint32_t qh1;

    uint64_t tmp0[4];
    uint64_t tmp1[4];

    for (; ib + 1 < nb; ib += 2) {
        const block_q5_0 * restrict x0 = &x[ib];
        const block_q5_0 * restrict x1 = &x[ib + 1];
        const block_q8_0 * restrict y0 = &y[ib];
        const block_q8_0 * restrict y1 = &y[ib + 1];

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

        // extract the 5th bit via lookup table ((!b) << 4)
        memcpy(&qh0, x0->qh, sizeof(qh0));
        memcpy(&qh1, x1->qh, sizeof(qh1));

        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];

        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];

        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));

        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
        const uint8x16_t v0_1 = vld1q_u8(x1->qs);

        // 4-bit -> 8-bit
        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));

        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);

        // load y
        const int8x16_t v1_0l = vld1q_s8(y0->qs);
        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
        const int8x16_t v1_1l = vld1q_s8(y1->qs);
        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__wasm_simd128__)
    v128_t sumv = wasm_f32x4_splat(0.0f);

    uint32_t qh;
    uint64_t tmp[4];

    // TODO: check if unrolling this is better
    for (; ib < nb; ++ib) {
        const block_q5_0 * restrict x0 = &x[ib];
        const block_q8_0 * restrict y0 = &y[ib];

        const v128_t m4b  = wasm_i8x16_splat(0x0F);

        // extract the 5th bit
        memcpy(&qh, x0->qh, sizeof(qh));

        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
        tmp[3] = table_b2b_1[(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);

        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
        const v128_t v0hf = wasm_i8x16_sub(v0h, 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);

        // 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(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
    }

    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#endif

    for (; ib < nb; ++ib) {
        uint32_t qh;
        memcpy(&qh, x[ib].qh, sizeof(qh));

        int sumi0 = 0;
        int sumi1 = 0;

        for (int j = 0; j < qk/2; ++j) {
            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));

            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);

            sumi0 += (x0 * y[ib].qs[j]);
            sumi1 += (x1 * y[ib].qs[j + qk/2]);
        }

        int sumi = sumi0 + sumi1;
        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
}

Here is the function that you need to convert:

void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;

    assert(n % qk == 0);
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
    UNUSED(bs);

    const block_q8_0 * restrict x = vx;
    const block_q8_0 * restrict y = vy;

    int ib = 0;
    float sumf = 0;

#if defined(__ARM_NEON)
    float32x4_t sumv0 = vdupq_n_f32(0.0f);
    float32x4_t sumv1 = vdupq_n_f32(0.0f);

    for (; ib + 1 < nb; ib += 2) {
        const block_q8_0 * restrict x0 = &x[ib + 0];
        const block_q8_0 * restrict x1 = &x[ib + 1];
        const block_q8_0 * restrict y0 = &y[ib + 0];
        const block_q8_0 * restrict y1 = &y[ib + 1];

        const int8x16_t x0_0 = vld1q_s8(x0->qs);
        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
        const int8x16_t x1_0 = vld1q_s8(x1->qs);
        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);

        // load y
        const int8x16_t y0_0 = vld1q_s8(y0->qs);
        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
        const int8x16_t y1_0 = vld1q_s8(y1->qs);
        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));

        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#endif
    for (; ib < nb; ++ib) {
        int sumi = 0;

        for (int j = 0; j < qk; j++) {
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
}

hjortholm commented Jan 29, 2025

Optimized prompt for converting code

hjortholm commented Jan 29, 2025

DeepSeek evaluation of the converted code.

The performance gains from converting the ggml_vec_dot_q8_0_q8_0 function to use WASM SIMD (versus a scalar fallback in WebAssembly) are expected to be significant, with potential improvements of 4–8× for the SIMD-optimized portion of the code. Here’s a breakdown of the factors and trade-offs:

Key Performance Drivers
SIMD Parallelism:

Each 128-bit WASM SIMD register processes 16 int8 elements at once, reducing the number of instructions needed for arithmetic.

The i32x4.dot_i16x8 instruction computes 4 accumulated 32-bit dot products in parallel, replacing 32 scalar operations per block of 16 elements.

Memory Efficiency:

Contiguous 128-bit loads (wasm_v128_load) align with SIMD-friendly memory access patterns, minimizing overhead compared to scalar byte-wise loads.

Reduced Loop Overhead:

The SIMD version processes 32 elements per iteration (two 128-bit registers for x and y), drastically reducing loop iterations compared to scalar code (which processes 1 element per iteration).

Comparison to ARM NEON
NEON (original code): Uses vdotq_s32 for direct 8-bit dot products and unrolls loops (2 blocks per iteration), achieving high throughput.

WASM SIMD (converted code):

Requires 8-bit → 16-bit extension (extend_low/high_i8x16) before using dot_i16x8, adding slight overhead.

Processes 1 block per iteration (vs. 2 in NEON), increasing loop overhead slightly.

Despite this, WASM SIMD still achieves close-to-NEON performance in practice for WebAssembly targets.

Expected Gains
Scalar Fallback (WebAssembly):

Without SIMD, scalar code requires 32 multiply-adds + loop overhead per block.

SIMD reduces this to ~4–6 vector instructions per block, yielding 4–8× speedups for the SIMD-optimized loop.

Real-World Impact:

For large tensors (common in LLMs), the SIMD-optimized loop will dominate runtime, leading to near-linear scaling with SIMD width.

Example: A 512-element dot product would see ~16× fewer loop iterations in SIMD vs. scalar.

Trade-Offs
Loop Unrolling:

The NEON version unrolls loops (2 blocks/iteration), but the WASM code processes 1 block/iteration. This slightly increases loop overhead but simplifies code maintenance.

8→16-Bit Extension:

WASM lacks direct 8-bit dot product instructions, requiring extension to 16-bit. This adds instructions but is still far faster than scalar.

Code-Specific Optimizations
The provided WASM SIMD code:

Uses wasm_f32x4_mul + wasm_f32x4_add to mimic NEON’s vmlaq_n_f32 (fused multiply-add).

Efficiently reduces partial sums with wasm_f32x4_extract_lane, minimizing horizontal addition overhead.

Conclusion
The WASM SIMD version should deliver 4–8× faster execution compared to the scalar fallback in WebAssembly, depending on runtime optimizations (e.g., V8 TurboFan). While slightly less efficient than ARM NEON due to 8→16-bit extension and loop unrolling differences, it brings WebAssembly performance much closer to native code for quantized LLM inference. For best results, ensure the runtime enables SIMD support (e.g., --experimental-wasm-simd in Node.js).

hjortholm commented Jan 29, 2025

DeepSeek evaluation of the converted code.

hjortholm commented Jan 29, 2025 •

edited

Loading

A second pass optimization prompt:
In your thinking, you state "Possible performance trade-offs: processing one block at a time may use fewer registers but may not utilize SIMD as efficiently as processing two blocks. However, given the example's approach, this is acceptable."
Do I read this correctly as there are further performance gains possible? If yes explore potential further performance gains and implement these

#elif defined(wasm_simd128)
v128_t sumv = wasm_f32x4_splat(0.0f);

// Process two blocks per iteration to match NEON's unrolling
for (; ib + 1 < nb; ib += 2) {
    const block_q8_0 * restrict x0 = &x[ib];
    const block_q8_0 * restrict x1 = &x[ib + 1];
    const block_q8_0 * restrict y0 = &y[ib];
    const block_q8_0 * restrict y1 = &y[ib + 1];

    // Load all 32 elements (two blocks) into SIMD registers
    const v128_t x0_0 = wasm_v128_load(x0->qs);
    const v128_t x0_1 = wasm_v128_load(x0->qs + 16);
    const v128_t x1_0 = wasm_v128_load(x1->qs);
    const v128_t x1_1 = wasm_v128_load(x1->qs + 16);

    const v128_t y0_0 = wasm_v128_load(y0->qs);
    const v128_t y0_1 = wasm_v128_load(y0->qs + 16);
    const v128_t y1_0 = wasm_v128_load(y1->qs);
    const v128_t y1_1 = wasm_v128_load(y1->qs + 16);

    // Helper function to compute dot product for a single 16-element pair
    #define DOT_PRODUCT(a, b) \
        wasm_i32x4_add( \
            wasm_i32x4_dot_i16x8( \
                wasm_i16x8_extend_low_i8x16(a), \
                wasm_i16x8_extend_low_i8x16(b) \
            ), \
            wasm_i32x4_dot_i16x8( \
                wasm_i16x8_extend_high_i8x16(a), \
                wasm_i16x8_extend_high_i8x16(b) \
            ) \
        )

    // Compute dot products for both blocks in parallel
    const v128_t dot0 = DOT_PRODUCT(x0_0, y0_0); // Block 0, first 16 elements
    const v128_t dot1 = DOT_PRODUCT(x0_1, y0_1); // Block 0, second 16 elements
    const v128_t dot2 = DOT_PRODUCT(x1_0, y1_0); // Block 1, first 16 elements
    const v128_t dot3 = DOT_PRODUCT(x1_1, y1_1); // Block 1, second 16 elements

    // Combine results and convert to float
    const v128_t sum01 = wasm_i32x4_add(wasm_i32x4_add(dot0, dot1), wasm_i32x4_add(dot2, dot3));
    const v128_t scales = wasm_f32x4_make(
        GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d),
        GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d),
        GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d),
        GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d)
    );

    // Accumulate scaled sums
    sumv = wasm_f32x4_add(sumv, 
        wasm_f32x4_mul(
            wasm_f32x4_convert_i32x4(sum01),
            scales
        )
    );
}

// Handle remaining block (if nb is odd)
for (; ib < nb; ++ib) {
    const block_q8_0 * restrict x0 = &x[ib];
    const block_q8_0 * restrict y0 = &y[ib];

    const v128_t x0_0 = wasm_v128_load(x0->qs);
    const v128_t x0_1 = wasm_v128_load(x0->qs + 16);
    const v128_t y0_0 = wasm_v128_load(y0->qs);
    const v128_t y0_1 = wasm_v128_load(y0->qs + 16);

    const v128_t dot0 = DOT_PRODUCT(x0_0, y0_0);
    const v128_t dot1 = DOT_PRODUCT(x0_1, y0_1);

    const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
    sumv = wasm_f32x4_add(sumv,
        wasm_f32x4_mul(
            wasm_f32x4_convert_i32x4(wasm_i32x4_add(dot0, dot1)),
            wasm_f32x4_splat(scale)
        )
    );
}

// Final reduction
sumf = wasm_f32x4_extract_lane(sumv, 0)
     + wasm_f32x4_extract_lane(sumv, 1)
     + wasm_f32x4_extract_lane(sumv, 2)
     + wasm_f32x4_extract_lane(sumv, 3);

#endif
``
Performance Analysis
Loop Unrolling:

Reduces loop overhead by processing 2 blocks/iteration (vs. 1), cutting branch checks by ~50%.

Matches NEON’s strategy for better instruction scheduling.

SIMD Throughput:

Processes 64 int8 elements (2 blocks × 32 elements) per iteration.

Uses i32x4.doti16x8 efficiently across all registers without register spilling.

Reduced Redundancy:

The DOT_PRODUCT macro avoids code duplication.

Combines scaled sums for both blocks into a single f32x4 multiplication, reducing conversions.

Expected Gains
~10–20% Speedup: By unrolling loops and reducing overhead, this optimized version closes the gap with NEON.

Better ILP: Modern CPUs can pipeline instructions more effectively when working on multiple independent blocks.

Trade-Offs
Code Complexity: Slightly harder to read but maintains the same algorithmic structure.

This implementation balances performance and readability, matching ARM NEON’s efficiency while adhering to WASM’s constraints.

hjortholm commented Jan 29, 2025

A second pass optimization prompt

hjortholm commented Jan 29, 2025

Prompt for measuring performance and further enhancement:

objective: Determine the total performance gains from newly introduced optimizations and explore potential improvements.

Performance Gains:
How much speedup or efficiency is expected compared to the previous implementation?
Which metrics (e.g., execution time, memory usage, throughput) should be measured, and how?
What tools or methods (e.g., profilers, benchmark suites) will validate these performance improvements? 
Optimization Angles:
Are there further optimizations in algorithmic design, concurrency, or hardware-specific tuning?
Could different compiler options, instruction sets, or data layouts enhance results?
What trade-offs (e.g., code complexity vs. performance) might arise?
Expected Outcomes:
Provide guidance on short-term improvements and long-term enhancements to maintain high performance.
Outline any potential bottlenecks or corner cases that may limit gains

Performance Gains Analysis
1. Expected Speedup vs. Previous Implementation
Previous (1 block/iteration): Processed 32 elements (1 block) per iteration, with loop overhead for each block.

Optimized (2 blocks/iteration):

Loop Unrolling: Reduces loop overhead by ~50% (fewer branch checks).

SIMD Utilization: Processes 64 elements (2 blocks) in parallel, leveraging instruction-level parallelism.

Expected Gains:

10–20% speedup over the initial WASM SIMD version (1 block/iteration).

4–6× faster than scalar WebAssembly code (non-SIMD fallback).

2. Key Metrics to Measure
Metric How to Measure Relevance
Execution Time Time per ggml_vec_dot_q8_0_q8_0 call using high-resolution timers (e.g., performance.now()). Direct measure of optimization impact.
Throughput Elements processed per second (e.g., (n / execution_time) * 1e6). Quantifies SIMD efficiency.
SIMD Utilization Assembly inspection (e.g., wasm-dis tool) to confirm SIMD instructions dominate. Ensures vectorization isn’t hindered by data dependencies or misalignment.
Memory Bandwidth Profiler metrics (e.g., cache misses, load/store throughput via perf stat in native runtimes). Identifies bottlenecks in data access.
3. Validation Tools & Methods
WebAssembly Tools:

Chrome DevTools: Profile WebAssembly execution with the Performance panel.

WABT (WebAssembly Binary Toolkit): Inspect generated SIMD instructions.

V8 Flags: Use --experimental-wasm-simd to enforce SIMD optimizations.

Benchmark Suites:

Microbenchmarks with varying n (e.g., 512, 1024, 2048 elements) to test scaling.

Cross-validate with ARM NEON timings to compare against native performance.

Profilers:

Linux Perf: If running in a WASM-to-native environment (e.g., Wasmer/Wasmtime).

Emscripten’s --profiling: Generate function-level timing reports.

Optimization Angles

Algorithmic Improvements
Reduced 8→16-Bit Extension Overhead:

Explore fused 8-bit dot product approximations (if WebAssembly adds i32x4.dot_i8x16 in future).

Block Size Tuning:

Experiment with larger blocks (e.g., QK8_0=64) to amortize loop overhead, but balance with cache locality.

2. Concurrency
Web Workers:

Split n into chunks processed in parallel across threads (limited by WebAssembly’s threading support).

SIMD + Multithreading:

Combine thread-level parallelism (e.g., SharedArrayBuffer) with SIMD for large tensors.

Hardware-Specific Tuning
Memory Alignment:

Ensure block_q8_0 structs are 16-byte aligned for faster v128.load operations.

CPU-Specific Scheduling:

Use compiler hints (e.g., #pragma unroll in Emscripten) to optimize for pipelining.

Compiler & Data Layout
Compiler Flags:

-msimd128 -O3 (Emscripten) to maximize SIMD optimizations.

-flto for link-time optimizations.

Data Layout:

Structure-of-Arrays (SoA) for qs/d fields to improve prefetching.

Trade-Offs
Optimization Benefit Trade-Off
Loop Unrolling Reduces branch overhead. Increases code size; harder to maintain.
SIMD Register Pressure Maximizes parallelism. Risk of spilling to memory on constrained hardware.
Threading Utilizes multiple cores. Adds complexity; WebAssembly threading is still experimental.
Expected Outcomes
Short-Term Improvements
Validate Current Optimizations:

Confirm 10–20% gains over the 1-block/iteration WASM SIMD code with microbenchmarks.

Memory Alignment:

Enforce 16-byte alignment for block_q8_0 to avoid unaligned loads.

Compiler Tuning:

Test -O3 vs. -Os to balance speed and code size.

Long-Term Enhancements
WebAssembly Future Features:

Adopt wider SIMD (e.g., 256-bit) if standardized.

Leverage i8x16.dot instructions if added to the spec.

Algorithmic Hybridization:

Mix SIMD and scalar code for small n (e.g., n < 256).

WebGPU Integration:

Offload large dot products to GPU compute shaders.

Bottlenecks & Corner Cases
Memory Bandwidth:

If tensors exceed L2 cache, SIMD gains may plateau. Use smaller blocks or tiling.

Odd Block Counts:

The fallback loop for residual blocks (ib < nb) adds minor overhead. Ensure it’s minimal.

Half-Precision Scaling:

GGML_FP16_TO_FP32 conversions are scalar; batch-convert d values upfront if possible.

Final Guidance
Immediate Next Steps:

Benchmark with real-world LLM inference workloads (e.g., a transformer layer).

Profile memory alignment impact using wasm-opt --align-features.

Future Roadmap:

Monitor WebAssembly SIMD evolution (e.g., relaxed SIMD proposals).

Explore WebGPU for heterogeneous compute in parallel with SIMD.

By balancing SIMD efficiency, loop unrolling, and memory optimizations, the current implementation achieves near-native performance for WebAssembly while leaving room for future gains as the ecosystem matures.

ngxson/FAQ.md

Why did you do this?

What is your setup?

Does it work in one-shot or I have to prompt it multiple times?

It only does conversion ARM NEON --> WASM SIMD, or it can invent new WASM SIMD code from scratch?

How much time does it spent to think?

Prompt is very long, what to do?

NOTE: This task (q6_K_q8_K) is HARD. The prompt below always produces failed result

ngxson commented Jan 27, 2025

Uh oh!

henryclw commented Jan 27, 2025

Uh oh!

timwu-ipevo commented Jan 27, 2025

Uh oh!

ngxson commented Jan 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Manamama commented Jan 28, 2025

Uh oh!

0xCA commented Jan 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ngxson commented Jan 28, 2025

Uh oh!

ngxson commented Jan 28, 2025

Uh oh!

0xCA commented Jan 28, 2025

Uh oh!

ngxson commented Jan 28, 2025

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

hjortholm commented Jan 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

hjortholm commented Jan 29, 2025

Uh oh!

ngxson commented Jan 27, 2025 •

edited

Loading

0xCA commented Jan 28, 2025 •

edited

Loading

hjortholm commented Jan 29, 2025 •

edited

Loading