diff --git a/ggml/src/ggml-hexagon/htp/act-ops.c b/ggml/src/ggml-hexagon/htp/act-ops.c index 7e488456ee..88bd2ddc43 100644 --- a/ggml/src/ggml-hexagon/htp/act-ops.c +++ b/ggml/src/ggml-hexagon/htp/act-ops.c @@ -85,13 +85,16 @@ static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0, struct htp_spad * dst_spad, uint32_t nth, uint32_t ith, - uint32_t src0_nrows_per_thread) { + uint32_t src0_nrows_per_thread, + dma_queue * dma_queue) { htp_act_preamble3; size_t src0_row_size = nb01; size_t src1_row_size = nb11; size_t dst_row_size = nb1; + + const uint32_t src0_nrows = ne01 * ne02 * ne03; // src0 rows const uint32_t src0_start_row = src0_nrows_per_thread * ith; @@ -105,10 +108,129 @@ static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0, uint64_t t1, t2; t1 = HAP_perf_get_qtimer_count(); - int is_aligned = 1; - if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) { - is_aligned = 0; - FARF(HIGH, "swiglu-f32: unaligned addresses in elementwise op, possibly slower execution\n"); + const uint8_t * restrict data_src0 = (const uint8_t *) src0->data; + const uint8_t * restrict data_src1 = (const uint8_t *) src1->data; + uint8_t * restrict data_dst = (uint8_t *) dst->data; + + const bool src1_valid = src1->ne[0]; + const int nc = (src1_valid) ? ne00 : ne00 / 2; + if (!src1_valid) { + const int32_t swapped = op_params[1]; + data_src1 = data_src0; + src1_row_size = src0_row_size; + + const size_t nc_in_bytes = nc * SIZEOF_FP32; + data_src0 += swapped ? nc_in_bytes : 0; + data_src1 += swapped ? 0 : nc_in_bytes; + } + + const size_t src0_row_size_aligned = htp_round_up(src0_row_size, VLEN); + const size_t src1_row_size_aligned = htp_round_up(src1_row_size, VLEN); + const size_t dst_row_size_aligned = htp_round_up(dst_row_size, VLEN); + + uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread); + uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread); + uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_spad->size_per_thread); + + // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0 + size_t src0_spad_half_size = src0_spad->size_per_thread / 2; + size_t src1_spad_half_size = src1_spad->size_per_thread / 2; + size_t dst_spad_half_size = dst_spad->size_per_thread / 2; + + const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block + if (BLOCK == 0) { + FARF(ERROR, + "swiglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n", + src0_spad->size_per_thread, src0_row_size_aligned); + return; + } + + // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379 + for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); + + // Dummy DMA transation for sequencing (interleaving dst,src,dst,...) + dma_queue_push_vtcm_to_ddr(dma_queue, + dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)), + dst_row_size, dst_row_size_aligned, 0); + + dma_queue_push_ddr_to_vtcm(dma_queue, + dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)), + src0_row_size_aligned, src0_row_size, block_size); + dma_queue_push_ddr_to_vtcm(dma_queue, + dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)), + src1_row_size_aligned, src1_row_size, block_size); + } + + for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); + + float * dst_spad = (float *) dma_queue_pop(dma_queue).src; + float * src0_spad = (float *) dma_queue_pop(dma_queue).dst; + float * src1_spad = (float *) dma_queue_pop(dma_queue).dst; + + for (uint32_t ib = 0; ib < block_size; ib++) { + const float * src0_spad_ptr = src0_spad + ib * (src0_row_size_aligned / sizeof(float)); + const float * src1_spad_ptr = src1_spad + ib * (src1_row_size_aligned / sizeof(float)); + float * dst_spad_ptr = dst_spad + ib * (dst_row_size_aligned / sizeof(float)); + + //swiglu(x) = x1 * sigmoid(x0) + hvx_fast_sigmoid_f32((const uint8_t *) src0_spad_ptr, (uint8_t *) dst_spad_ptr, nc); + hvx_mul_mul_f32_opt((const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, + (const uint8_t *) src1_spad_ptr, (uint8_t *) dst_spad_ptr, nc); + } + + dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size, + dst_row_size_aligned, block_size); + + // prefetch N+2 loop iteration if any + const uint32_t pref_block = (ir + BLOCK * 2); + if (pref_block < src0_end_row) { + const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block); + dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)), + src0_row_size_aligned, src0_row_size, pref_block_size); + dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)), + src1_row_size_aligned, src1_row_size, pref_block_size); + } + } + + dma_queue_flush(dma_queue); + + t2 = HAP_perf_get_qtimer_count(); + + FARF(HIGH, "swiglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, + ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, + (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1)); +} + +static void glu_swiglu_oai_fp32_per_thread(const struct htp_tensor * src0, + const struct htp_tensor * src1, + struct htp_tensor * dst, + const int32_t * op_params, + struct htp_spad * src0_spad, + struct htp_spad * src1_spad, + struct htp_spad * dst_spad, + uint32_t nth, + uint32_t ith, + uint32_t src0_nrows_per_thread, + dma_queue * dma_queue) { + htp_act_preamble3; + + uint64_t t1, t2; + t1 = HAP_perf_get_qtimer_count(); + + size_t src0_row_size = nb01; + size_t src1_row_size = nb11; + size_t dst_row_size = nb1; + + const uint32_t src0_nrows = ne01 * ne02 * ne03; // src0 rows + + const uint32_t src0_start_row = src0_nrows_per_thread * ith; + const uint32_t src0_end_row = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows); + + // no work for this thread + if (src0_start_row >= src0_end_row) { + return; } const uint8_t * restrict data_src0 = (const uint8_t *) src0->data; @@ -127,130 +249,94 @@ static void glu_swiglu_fp32_per_thread(const struct htp_tensor * src0, data_src1 += swapped ? 0 : nc_in_bytes; } - uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size); - uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_row_size); - uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_row_size); + const size_t src0_row_size_aligned = htp_round_up(src0_row_size, VLEN); + const size_t src1_row_size_aligned = htp_round_up(src1_row_size, VLEN); + const size_t dst_row_size_aligned = htp_round_up(dst_row_size, VLEN); - const bool opt_path = ((1 == is_aligned) && !(nb01 & (VLEN - 1))); - for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) { - const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size)); - const float * restrict src1 = (float *) (data_src1 + (ir * src1_row_size)); - float * restrict dst = (float *) (data_dst + (ir * dst_row_size)); + uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread); + uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread); + uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_spad->size_per_thread); - if (ir + 1 < src0_end_row) { - htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size); - } + // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0 + size_t src0_spad_half_size = src0_spad->size_per_thread / 2; + size_t src1_spad_half_size = src1_spad->size_per_thread / 2; + size_t dst_spad_half_size = dst_spad->size_per_thread / 2; - if (opt_path) { - hvx_fast_sigmoid_f32((const uint8_t *) src0, (uint8_t *) src0_spad_data, nc); - hvx_mul_mul_f32_opt((const uint8_t *) src0, (const uint8_t *) src0_spad_data, (const uint8_t *) src1, - (uint8_t *) dst, nc); - } else { - hvx_exp_f32((const uint8_t *) src0, src0_spad_data, nc, true); - hvx_add_scalar_f32(src0_spad_data, 1.0, src1_spad_data, nc); - hvx_inverse_f32(src1_spad_data, src0_spad_data, nc); - - hvx_mul_f32((const uint8_t *) src0, src0_spad_data, dst_spad_data, nc); - hvx_mul_f32(dst_spad_data, (const uint8_t *) src1, (uint8_t *) dst, nc); - } - } - - t2 = HAP_perf_get_qtimer_count(); - - FARF(HIGH, "swiglu-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, - ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, - (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1)); -} - -static void glu_swiglu_oai_fp32_per_thread(const struct htp_tensor * src0, - const struct htp_tensor * src1, - struct htp_tensor * dst, - const int32_t * op_params, - struct htp_spad * src0_spad, - struct htp_spad * src1_spad, - struct htp_spad * dst_spad, - uint32_t nth, - uint32_t ith, - uint32_t src0_nrows_per_thread) { - htp_act_preamble3; - - uint64_t t1, t2; - t1 = HAP_perf_get_qtimer_count(); - - const size_t src0_row_size = nb01; - const size_t src1_row_size = nb11; - const size_t dst_row_size = nb1; - - const uint32_t src0_nrows = ne01 * ne02 * ne03; // src0 rows - - const uint32_t src0_start_row = src0_nrows_per_thread * ith; - const uint32_t src0_end_row = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows); - - // no work for this thread - if (src0_start_row >= src0_end_row) { + const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block + if (BLOCK == 0) { + FARF(ERROR, + "swiglu-oai-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least " + "%zu\n", + src0_spad->size_per_thread, src0_row_size_aligned); return; } + const float alpha = ((const float *) (op_params))[2]; + const float limit = ((const float *) (op_params))[3]; - if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) { - FARF(HIGH, "act-f32: unaligned addresses in activations op, possibly slower execution\n"); + // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379 + for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); + + // Dummy DMA transation for sequencing (interleaving dst,src,dst,...) + dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)), + dst_row_size, dst_row_size_aligned, 0); + + dma_queue_push_ddr_to_vtcm( + dma_queue, + dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)), + src0_row_size_aligned, src0_row_size, block_size); + dma_queue_push_ddr_to_vtcm( + dma_queue, + dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)), + src1_row_size_aligned, src1_row_size, block_size); } - const uint8_t * restrict data_src0 = (const uint8_t *) src0->data; - const uint8_t * restrict data_src1 = (const uint8_t *) src1->data; - uint8_t * restrict data_dst = (uint8_t *) dst->data; + for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); - bool src1_valid = src1->ne[0]; - if (!src1_valid) { - data_src1 = data_src0; - } + float * dst_spad = (float *) dma_queue_pop(dma_queue).src; + float * src0_spad = (float *) dma_queue_pop(dma_queue).dst; + float * src1_spad = (float *) dma_queue_pop(dma_queue).dst; - uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size); - uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_row_size); - uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_row_size); + for (uint32_t ib = 0; ib < block_size; ib++) { + const float * src0_spad_ptr = src0_spad + ib * (src0_row_size_aligned / sizeof(float)); + const float * src1_spad_ptr = src1_spad + ib * (src1_row_size_aligned / sizeof(float)); + float * dst_spad_ptr = dst_spad + ib * (dst_row_size_aligned / sizeof(float)); - const int32_t swapped = op_params[1]; - const float alpha = ((const float *) (op_params))[2]; - const float limit = ((const float *) (op_params))[3]; - - const int nc = (src1_valid) ? ne00 : ne00 / 2; - - for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) { - const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size)); - const float * restrict src1 = (float *) (data_src1 + (ir * src1_row_size)); - float * restrict dst = (float *) (data_dst + (ir * dst_row_size)); - - if (ir + 1 < src0_end_row) { - htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size); + // x (src0_spad_data) = std::min(src0_p[k], limit); + hvx_min_scalar_f32((const uint8_t *) src0_spad_ptr, limit, (uint8_t *) src0_spad_ptr, nc); + // y1 (src1_spad_data) = std::clamp(src1_p[k], -limit, limit); + hvx_clamp_scalar_f32((const uint8_t *) src1_spad_ptr, -limit, limit, (uint8_t *) src1_spad_ptr, nc); + // y (src1_spad_data) = y1 + 1.f + hvx_add_scalar_f32((const uint8_t *) src1_spad_ptr, 1.0, (uint8_t *) src1_spad_ptr, nc); + // x1 (dst_spad_data) = alpha * (x) + hvx_mul_scalar_f32((const uint8_t *) src0_spad_ptr, alpha, (uint8_t *) dst_spad_ptr, nc); + // x2 (dst_spad_data) = sigmoid(x1) = 1/(1+exp(-x1)) + hvx_fast_sigmoid_f32((const uint8_t *) dst_spad_ptr, (uint8_t *) dst_spad_ptr, nc); + // out = x * sigmoid(alpha * x) * (y + 1.f) + hvx_mul_mul_f32_opt((const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, + (const uint8_t *) src1_spad_ptr, (uint8_t *) dst_spad_ptr, nc); } - if (!src1) { - src0 += swapped ? nc : 0; - src1 += swapped ? 0 : nc; - } + dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size, + dst_row_size_aligned, block_size); - // x (src0_spad_data) = std::min(src0_p[k], limit); - hvx_min_scalar_f32((const uint8_t *) src0, limit, src0_spad_data, nc); - // y1 (src1_spad_data) = std::clamp(src1_p[k], -limit, limit); - hvx_clamp_scalar_f32((const uint8_t *) src1, -limit, limit, src1_spad_data, nc); - // y (src1_spad_data) = y1 + 1.f - hvx_add_scalar_f32(src1_spad_data, 1.0, src1_spad_data, nc); - // x1 (dst_spad_data) = alpha * (x) - hvx_mul_scalar_f32(src0_spad_data, alpha, dst_spad_data, nc); - // x2 (dst_spad_data) = expf(-x1) - hvx_exp_f32(dst_spad_data, dst_spad_data, nc, true); - // x3 (dst_spad_data) = x2 + 1.f - hvx_add_scalar_f32(dst_spad_data, 1.0, dst_spad_data, nc); - // x4 (dst_spad_data) = 1 / x3 - hvx_inverse_f32(dst_spad_data, dst_spad_data, nc); - // out_glu(dst_spad_data) = x * x4 - hvx_mul_f32(src0_spad_data, dst_spad_data, dst_spad_data, nc); - // out = out_glu * (y + 1.f); - hvx_mul_f32(dst_spad_data, src1_spad_data, (uint8_t *) dst, nc); + // prefetch N+2 loop iteration if any + const uint32_t pref_block = (ir + BLOCK * 2); + if (pref_block < src0_end_row) { + const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block); + dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)), + src0_row_size_aligned, src0_row_size, pref_block_size); + dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)), + src1_row_size_aligned, src1_row_size, pref_block_size); + } } + dma_queue_flush(dma_queue); + t2 = HAP_perf_get_qtimer_count(); - FARF(HIGH, "swiglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, src0->ne[0], + FARF(HIGH, "swiglu-oai-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src0_start_row, src0_end_row, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1)); } @@ -371,7 +457,8 @@ static void unary_silu_fp32_per_thread(const struct htp_tensor * src0, struct htp_spad * dst_spad, uint32_t nth, uint32_t ith, - uint32_t src0_nrows_per_thread) { + uint32_t src0_nrows_per_thread, + dma_queue * dma_queue) { htp_act_preamble2; uint64_t t1, t2; @@ -379,6 +466,8 @@ static void unary_silu_fp32_per_thread(const struct htp_tensor * src0, const size_t src0_row_size = nb01; const size_t dst_row_size = nb1; + const size_t src0_row_size_aligned = htp_round_up(src0_row_size, VLEN); + const size_t dst_row_size_aligned = htp_round_up(dst_row_size, VLEN); const uint32_t src0_nrows = ne01 * ne02 * ne03; @@ -390,64 +479,91 @@ static void unary_silu_fp32_per_thread(const struct htp_tensor * src0, return; } - int is_aligned = 1; - int opt_path = 0; - if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) { - is_aligned = 0; - FARF(HIGH, "silu-f32: unaligned addresses in elementwise op, possibly slower execution\n"); - } - if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) { - opt_path = 1; + const uint8_t * data_src0 = (const uint8_t *) src0->data; + uint8_t * data_dst = (uint8_t *) dst->data; + + uint8_t * src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread); + uint8_t * dst_spad_data = dst_spad->data + (ith * dst_spad->size_per_thread); + + // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0 + size_t src0_spad_half_size = src0_spad->size_per_thread / 2; + size_t dst_spad_half_size = dst_spad->size_per_thread / 2; + + const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block + + if (BLOCK == 0) { + FARF(ERROR, "silu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n", + src0_spad->size_per_thread, src0_row_size_aligned); + return; } - const uint8_t * restrict data_src0 = (const uint8_t *) src0->data; - uint8_t * restrict data_dst = (uint8_t *) dst->data; + // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379 + for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); - uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size); - uint8_t * restrict dst_spad_data = dst_spad->data + (ith * dst_row_size); + // Dummy DMA transation for sequencing (interleaving dst,src,dst,...) + dma_queue_push_vtcm_to_ddr(dma_queue, + dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)), + dst_row_size, dst_row_size_aligned, 0); - for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) { - const float * restrict src0 = (float *) (data_src0 + (ir * src0_row_size)); - float * restrict dst = (float *) (data_dst + (ir * dst_row_size)); + dma_queue_push_ddr_to_vtcm(dma_queue, + dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)), + src0_row_size_aligned, src0_row_size, block_size); + } - if (ir + 1 < src0_end_row) { - htp_l2fetch(src0 + src0_row_size, 1, src0_row_size, src0_row_size); + for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) { + const uint32_t block_size = MIN(BLOCK, src0_end_row - ir); + + float* dst_spad = (float *) dma_queue_pop(dma_queue).src; + float* src0_spad = (float *) dma_queue_pop(dma_queue).dst; + + for (uint32_t ib = 0; ib < block_size; ib++) { + const float* src0_spad_ptr = src0_spad + ib * (src0_row_size_aligned / sizeof(float)); + float* dst_spad_ptr = dst_spad + ib * (dst_row_size_aligned / sizeof(float)); + + // silu = x * sigmoid(x) + hvx_fast_sigmoid_f32((const uint8_t *) src0_spad_ptr, (uint8_t *) dst_spad_ptr, ne0); + hvx_mul_f32_opt((const uint8_t *) src0_spad_ptr, (uint8_t *) dst_spad_ptr, (uint8_t *) dst_spad_ptr, ne0); } - if (1 == opt_path) { - hvx_fast_sigmoid_f32((const uint8_t *) src0, (uint8_t *) src0_spad_data, ne0); - hvx_mul_f32_opt((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0); - } else { - hvx_exp_f32((const uint8_t *) src0, src0_spad_data, ne0, true); - hvx_add_scalar_f32(src0_spad_data, 1.0, dst_spad_data, ne0); - hvx_inverse_f32(dst_spad_data, src0_spad_data, ne0); + dma_queue_push_vtcm_to_ddr(dma_queue, + dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), + dst_row_size, dst_row_size_aligned, block_size); - hvx_mul_f32((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0); + // prefetch N+2 loop iteration if any + const uint32_t pref_block = (ir + BLOCK * 2); + if (pref_block < src0_end_row) { + const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block); + dma_queue_push_ddr_to_vtcm(dma_queue, + dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)), + src0_row_size_aligned, src0_row_size, pref_block_size); } } + dma_queue_flush(dma_queue); + t2 = HAP_perf_get_qtimer_count(); - FARF(HIGH, "silu-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, ne00, ne01, ne02, + FARF(HIGH, "silu-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1)); } static void unary_silu_fp32(unsigned int n, unsigned int i, void * data) { struct htp_ops_context * octx = (struct htp_ops_context *) data; unary_silu_fp32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i, - octx->src0_nrows_per_thread); + octx->src0_nrows_per_thread, octx->ctx->dma[i]); } static void glu_swiglu_fp32(unsigned int n, unsigned int i, void * data) { struct htp_ops_context * octx = (struct htp_ops_context *) data; glu_swiglu_fp32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad, - &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread); + &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]); } static void glu_swiglu_oai_fp32(unsigned int n, unsigned int i, void * data) { struct htp_ops_context * octx = (struct htp_ops_context *) data; glu_swiglu_oai_fp32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad, - &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread); + &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]); } static int execute_op_activations_fp32(struct htp_ops_context * octx) {