From 6ff12a0f7765f62b8d0fa8554021e1cac2789f19 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Thu, 2 Nov 2017 16:09:35 +0000 Subject: COMPMID-662: Integrated the new a64_s8_gemm_12x8 + dot product kernel into ACL. Change-Id: Id8f919e486a132fc58346c9f84fccbeeb83d19b3 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94233 Tested-by: Kaizen Reviewed-by: Anthony Barbier Reviewed-by: Gian Marco Iodice --- src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp | 38 +- .../kernels/NEGEMMLowpMatrixMultiplyKernel.cpp | 130 +++--- .../NEON/kernels/NEGEMMLowpReductionKernel.cpp | 174 ++++---- .../kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp | 483 ++------------------- 4 files changed, 218 insertions(+), 607 deletions(-) (limited to 'src/core/NEON/kernels') diff --git a/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp index 400c6d9d8c..255e486365 100644 --- a/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp @@ -145,7 +145,7 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32); // Convert S32 to U16 - const int16x8x2_t in_u16 = + const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), @@ -153,10 +153,10 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) } }; - // Convert U16 to U8 - const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_u16.val[0]), vqmovun_s16(in_u16.val[1])); + // Convert S16 to S8 + const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); - vst1q_u8(out.ptr(), out_u8); + vst1q_s8(reinterpret_cast(out.ptr()), out_s8); }, vector_sum_col, vector_sum_row, mm_result, out); } @@ -209,7 +209,7 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32); // Convert S32 to U16 - const int16x8x2_t in_u16 = + const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), @@ -217,10 +217,10 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) } }; - // Convert U16 to U8 - const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_u16.val[0]), vqmovun_s16(in_u16.val[1])); + // Convert S16 to S8 + const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); - vst1q_u8(out.ptr(), out_u8); + vst1q_s8(reinterpret_cast(out.ptr()), out_s8); }, vector_sum_row, mm_result, out); } @@ -295,8 +295,8 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32); in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32); - // Convert S32 to U16 - const int16x8x2_t in_u16 = + // Convert S32 to S16 + const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), @@ -304,10 +304,10 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) } }; - // Convert U16 to U8 - const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_u16.val[0]), vqmovun_s16(in_u16.val[1])); + // Convert S16 to S8 + const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); - vst1q_u8(out.ptr(), out_u8); + vst1q_s8(reinterpret_cast(out.ptr()), out_s8); }, vector_sum_col, mm_result, out); } @@ -346,8 +346,8 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32); in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32); - // Convert S32 to U16 - const int16x8x2_t in_u16 = + // Convert S32 to S16 + const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), @@ -355,10 +355,10 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window) } }; - // Convert U16 to U8 - const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_u16.val[0]), vqmovun_s16(in_u16.val[1])); + // Convert U16 to S8 + const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); - vst1q_u8(out.ptr(), out_u8); + vst1q_s8(reinterpret_cast(out.ptr()), out_s8); }, mm_result, out); } @@ -375,7 +375,7 @@ void NEGEMMLowpFinalizeKernel::configure(const ITensor *vector_sum_col, const IT int32_t c_offset, int32_t c_mult_int, int32_t shift) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S8); TensorShape mm_result_shape = mm_result->info()->tensor_shape(); TensorShape output_shape = output->info()->tensor_shape(); diff --git a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp index 3e614a8bfc..4b9c9f3e64 100644 --- a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp @@ -52,7 +52,7 @@ NEGEMMLowpMatrixMultiplyKernel::NEGEMMLowpMatrixMultiplyKernel() void NEGEMMLowpMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::U8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); @@ -127,115 +127,115 @@ void NEGEMMLowpMatrixMultiplyKernel::run(const Window &window, const ThreadInfo // All the values needed for computing a single 4x4 block will be read from consecutive memory positions execute_window_loop(window, [&](const Coordinates & id) { - const uint8_t *mtx_a0 = ina.ptr(); - const uint8_t *mtx_b0 = inb.ptr(); + auto *mtx_a0 = reinterpret_cast(ina.ptr()); + auto *mtx_b0 = reinterpret_cast(inb.ptr()); // Note: Since the input are all positives, we can use uint32_t // Accumulators for the block 0 - uint32x4x4_t c0 = + int32x4x4_t c0 = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; // Accumulators for the block 1 - uint32x4x4_t c1 = + int32x4x4_t c1 = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; // Accumulators for the block 2 - uint32x4x4_t c2 = + int32x4x4_t c2 = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; // Accumulators for the block 3 - uint32x4x4_t c3 = + int32x4x4_t c3 = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; for(int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16) { - const uint8x8_t a00_u8 = vld1_u8(mtx_a0); - const uint8x16_t b00_u8 = vld1q_u8(mtx_b0); + const int8x8_t a00_s8 = vld1_s8(mtx_a0); + const int8x16_t b00_s8 = vld1q_s8(mtx_b0); - // Convert a00_u8 to uint16_t and get the lower part - const uint16x4_t a00_u16 = vget_low_u16(vmovl_u8(a00_u8)); + // Convert a00_s8 to uint16_t and get the lower part + const int16x4_t a00_s16 = vget_low_s16(vmovl_s8(a00_s8)); - // Convert b00_u8 to int16_t - const uint16x4x4_t b00_u16 = + // Convert b00_s8 to int16_t + const int16x4x4_t b00_s16 = { { - vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))), - vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))), - vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))), - vget_high_u16(vmovl_u8(vget_high_u8(b00_u8))) + vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))), + vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))), + vget_high_s16(vmovl_s8(vget_high_s8(b00_s8))) } }; // 4x4 block 0 - c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16, 0); - c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16, 0); - c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16, 0); - c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16, 0); + c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16, 0); + c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16, 0); + c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16, 0); + c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16, 0); // 4x4 block 1 - c1.val[0] = vmlal_lane_u16(c1.val[0], b00_u16.val[0], a00_u16, 1); - c1.val[1] = vmlal_lane_u16(c1.val[1], b00_u16.val[1], a00_u16, 1); - c1.val[2] = vmlal_lane_u16(c1.val[2], b00_u16.val[2], a00_u16, 1); - c1.val[3] = vmlal_lane_u16(c1.val[3], b00_u16.val[3], a00_u16, 1); + c1.val[0] = vmlal_lane_s16(c1.val[0], b00_s16.val[0], a00_s16, 1); + c1.val[1] = vmlal_lane_s16(c1.val[1], b00_s16.val[1], a00_s16, 1); + c1.val[2] = vmlal_lane_s16(c1.val[2], b00_s16.val[2], a00_s16, 1); + c1.val[3] = vmlal_lane_s16(c1.val[3], b00_s16.val[3], a00_s16, 1); // 4x4 block 2 - c2.val[0] = vmlal_lane_u16(c2.val[0], b00_u16.val[0], a00_u16, 2); - c2.val[1] = vmlal_lane_u16(c2.val[1], b00_u16.val[1], a00_u16, 2); - c2.val[2] = vmlal_lane_u16(c2.val[2], b00_u16.val[2], a00_u16, 2); - c2.val[3] = vmlal_lane_u16(c2.val[3], b00_u16.val[3], a00_u16, 2); + c2.val[0] = vmlal_lane_s16(c2.val[0], b00_s16.val[0], a00_s16, 2); + c2.val[1] = vmlal_lane_s16(c2.val[1], b00_s16.val[1], a00_s16, 2); + c2.val[2] = vmlal_lane_s16(c2.val[2], b00_s16.val[2], a00_s16, 2); + c2.val[3] = vmlal_lane_s16(c2.val[3], b00_s16.val[3], a00_s16, 2); // 4x4 block 3 - c3.val[0] = vmlal_lane_u16(c3.val[0], b00_u16.val[0], a00_u16, 3); - c3.val[1] = vmlal_lane_u16(c3.val[1], b00_u16.val[1], a00_u16, 3); - c3.val[2] = vmlal_lane_u16(c3.val[2], b00_u16.val[2], a00_u16, 3); - c3.val[3] = vmlal_lane_u16(c3.val[3], b00_u16.val[3], a00_u16, 3); + c3.val[0] = vmlal_lane_s16(c3.val[0], b00_s16.val[0], a00_s16, 3); + c3.val[1] = vmlal_lane_s16(c3.val[1], b00_s16.val[1], a00_s16, 3); + c3.val[2] = vmlal_lane_s16(c3.val[2], b00_s16.val[2], a00_s16, 3); + c3.val[3] = vmlal_lane_s16(c3.val[3], b00_s16.val[3], a00_s16, 3); } auto mtx_out = reinterpret_cast(out.ptr()); - vst1q_s32(mtx_out + 0 * out_stride + 0, vreinterpretq_s32_u32(c0.val[0])); - vst1q_s32(mtx_out + 0 * out_stride + 4, vreinterpretq_s32_u32(c0.val[1])); - vst1q_s32(mtx_out + 0 * out_stride + 8, vreinterpretq_s32_u32(c0.val[2])); - vst1q_s32(mtx_out + 0 * out_stride + 12, vreinterpretq_s32_u32(c0.val[3])); - vst1q_s32(mtx_out + 1 * out_stride + 0, vreinterpretq_s32_u32(c1.val[0])); - vst1q_s32(mtx_out + 1 * out_stride + 4, vreinterpretq_s32_u32(c1.val[1])); - vst1q_s32(mtx_out + 1 * out_stride + 8, vreinterpretq_s32_u32(c1.val[2])); - vst1q_s32(mtx_out + 1 * out_stride + 12, vreinterpretq_s32_u32(c1.val[3])); - vst1q_s32(mtx_out + 2 * out_stride + 0, vreinterpretq_s32_u32(c2.val[0])); - vst1q_s32(mtx_out + 2 * out_stride + 4, vreinterpretq_s32_u32(c2.val[1])); - vst1q_s32(mtx_out + 2 * out_stride + 8, vreinterpretq_s32_u32(c2.val[2])); - vst1q_s32(mtx_out + 2 * out_stride + 12, vreinterpretq_s32_u32(c2.val[3])); - vst1q_s32(mtx_out + 3 * out_stride + 0, vreinterpretq_s32_u32(c3.val[0])); - vst1q_s32(mtx_out + 3 * out_stride + 4, vreinterpretq_s32_u32(c3.val[1])); - vst1q_s32(mtx_out + 3 * out_stride + 8, vreinterpretq_s32_u32(c3.val[2])); - vst1q_s32(mtx_out + 3 * out_stride + 12, vreinterpretq_s32_u32(c3.val[3])); + vst1q_s32(mtx_out + 0 * out_stride + 0, c0.val[0]); + vst1q_s32(mtx_out + 0 * out_stride + 4, c0.val[1]); + vst1q_s32(mtx_out + 0 * out_stride + 8, c0.val[2]); + vst1q_s32(mtx_out + 0 * out_stride + 12, c0.val[3]); + vst1q_s32(mtx_out + 1 * out_stride + 0, c1.val[0]); + vst1q_s32(mtx_out + 1 * out_stride + 4, c1.val[1]); + vst1q_s32(mtx_out + 1 * out_stride + 8, c1.val[2]); + vst1q_s32(mtx_out + 1 * out_stride + 12, c1.val[3]); + vst1q_s32(mtx_out + 2 * out_stride + 0, c2.val[0]); + vst1q_s32(mtx_out + 2 * out_stride + 4, c2.val[1]); + vst1q_s32(mtx_out + 2 * out_stride + 8, c2.val[2]); + vst1q_s32(mtx_out + 2 * out_stride + 12, c2.val[3]); + vst1q_s32(mtx_out + 3 * out_stride + 0, c3.val[0]); + vst1q_s32(mtx_out + 3 * out_stride + 4, c3.val[1]); + vst1q_s32(mtx_out + 3 * out_stride + 8, c3.val[2]); + vst1q_s32(mtx_out + 3 * out_stride + 12, c3.val[3]); }, ina, inb, out); } diff --git a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp index 3f841bbf59..9df13ce0e3 100644 --- a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp @@ -51,7 +51,7 @@ INEGEMMLowpReductionKernel::INEGEMMLowpReductionKernel() void NEGEMMLowpMatrixAReductionKernel::configure(const ITensor *mtx_a_interleaved4x4, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a_interleaved4x4, 1, DataType::U8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a_interleaved4x4, 1, DataType::S8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); _input = mtx_a_interleaved4x4; @@ -97,9 +97,9 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf execute_window_loop(collapsed_window, [&](const Coordinates & id) { // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation - uint32x4_t sum_row = vdupq_n_u32(0); + int32x4_t sum_row = vdupq_n_s32(0); - const uint8_t *matrix_a = in.ptr() + (id.x() / 4) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]; + auto matrix_a = reinterpret_cast(in.ptr() + (id.x() / 4) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]); #if __arm__ asm volatile("PLD [%0, #128*4]" ::"r"(matrix_a)); @@ -109,43 +109,43 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf // This for loop performs 4 accumulations for(; i <= (_k - 4); i += 4) { - const uint8x16_t a0_u8 = vld1q_u8(matrix_a + i * 4); + const int8x16_t a0_s8 = vld1q_s8(matrix_a + i * 4); // Convert U8 to U16 - uint16x4x4_t a0_u16 = + int16x4x4_t a0_s16 = { { - vget_low_u16(vmovl_u8(vget_low_u8(a0_u8))), - vget_high_u16(vmovl_u8(vget_low_u8(a0_u8))), - vget_low_u16(vmovl_u8(vget_high_u8(a0_u8))), - vget_high_u16(vmovl_u8(vget_high_u8(a0_u8))) + vget_low_s16(vmovl_s8(vget_low_s8(a0_s8))), + vget_high_s16(vmovl_s8(vget_low_s8(a0_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(a0_s8))), + vget_high_s16(vmovl_s8(vget_high_s8(a0_s8))) } }; // Accumulate to U16 - a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[1]); - a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[2]); - a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[3]); + a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[1]); + a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[2]); + a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[3]); // Accumulate to U32 - sum_row = vaddw_u16(sum_row, a0_u16.val[0]); + sum_row = vaddw_s16(sum_row, a0_s16.val[0]); } // This for loop performs the leftover accumulations for(; i < _k; ++i) { - const uint8x8_t a0_u8 = vld1_u8(matrix_a + i * 4); + const int8x8_t a0_s8 = vld1_s8(matrix_a + i * 4); // Convert U8 to U16 - const uint16x4_t a0_u16 = vget_low_u16(vmovl_u8(a0_u8)); + const int16x4_t a0_s16 = vget_low_s16(vmovl_s8(a0_s8)); // Accumulate to U32 - sum_row = vaddw_u16(sum_row, a0_u16); + sum_row = vaddw_s16(sum_row, a0_s16); } auto vector_sum_row = reinterpret_cast(out.ptr()); - vst1q_s32(vector_sum_row, vreinterpretq_s32_u32(sum_row)); + vst1q_s32(vector_sum_row, sum_row); }, in, out); } @@ -154,10 +154,10 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf execute_window_loop(collapsed_window, [&](const Coordinates & id) { // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation - uint32x4_t sum_row_s32 = vdupq_n_u32(0); - unsigned int sum_row = 0; + int32x4_t sum_row_s32 = vdupq_n_s32(0); + int32_t sum_row = 0; - const uint8_t *matrix_a = in.ptr() + id.x() * _input->info()->strides_in_bytes()[1] + +id.y() * _input->info()->strides_in_bytes()[2]; + auto matrix_a = reinterpret_cast(in.ptr() + id.x() * _input->info()->strides_in_bytes()[1] + +id.y() * _input->info()->strides_in_bytes()[2]); #if __arm__ asm volatile("PLD [%0, #128*4]" ::"r"(matrix_a)); @@ -167,29 +167,29 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf // This for loop performs 16 accumulations for(; i <= (_k - 16); i += 16) { - const uint8x16_t a0_u8 = vld1q_u8(matrix_a + i); + const int8x16_t a0_s8 = vld1q_s8(matrix_a + i); // Partial accumulations in U16 - const uint16x8_t tmp_sum0 = vaddl_u8(vget_low_u8(a0_u8), vget_high_u8(a0_u8)); + const int16x8_t tmp_sum0 = vaddl_s8(vget_low_s8(a0_s8), vget_high_s8(a0_s8)); // Accumulate to U32 - sum_row_s32 = vaddq_u32(sum_row_s32, vpaddlq_u16(tmp_sum0)); + sum_row_s32 = vaddq_s32(sum_row_s32, vpaddlq_s16(tmp_sum0)); } // This for loop performs the leftover accumulations for(; i < _k; ++i) { - sum_row += static_cast(matrix_a[i]); + sum_row += static_cast(matrix_a[i]); } #if defined(__aarch64__) // Reduction operation available on 64 bit architectures only - sum_row += vaddvq_u32(sum_row_s32); + sum_row += vaddvq_s32(sum_row_s32); #else // __aarch64__ - uint32x2_t tmp = vpadd_u32(vget_high_u32(sum_row_s32), vget_low_u32(sum_row_s32)); - tmp = vpadd_u32(tmp, tmp); + int32x2_t tmp = vpadd_s32(vget_high_s32(sum_row_s32), vget_low_s32(sum_row_s32)); + tmp = vpadd_s32(tmp, tmp); - sum_row += vget_lane_u32(tmp, 0); + sum_row += vget_lane_s32(tmp, 0); #endif // __aarch64__ *(reinterpret_cast(out.ptr())) = static_cast(sum_row); @@ -200,7 +200,7 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf void NEGEMMLowpMatrixBReductionKernel::configure(const ITensor *mtx_b_transposed1xW, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b_transposed1xW, 1, DataType::U8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b_transposed1xW, 1, DataType::S8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); _input = mtx_b_transposed1xW; @@ -246,17 +246,17 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf execute_window_loop(collapsed_window, [&](const Coordinates & id) { // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation - uint32x4x4_t sum_col = + int32x4x4_t sum_col = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; - const uint8_t *matrix_b = in.ptr() + (id.x() / 16) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]; + auto matrix_b = reinterpret_cast(in.ptr() + (id.x() / 16) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]); #if __arm__ asm volatile("PLD [%0, #128*4]" ::"r"(matrix_b)); @@ -265,14 +265,14 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf int i = 0; for(; i < _k; ++i) { - const uint8x16_t b0_u8 = vld1q_u8(matrix_b + i * 16); + const int8x16_t b0_s8 = vld1q_s8(matrix_b + i * 16); - // Convert U8 to U16 - const uint16x8x2_t b0_u16 = + // Convert S8 to U16 + const int16x8x2_t b0_s16 = { { - vmovl_u8(vget_low_u8(b0_u8)), - vmovl_u8(vget_high_u8(b0_u8)) + vmovl_s8(vget_low_s8(b0_s8)), + vmovl_s8(vget_high_s8(b0_s8)) } }; @@ -280,20 +280,20 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf sum_col = { { - vaddw_u16(sum_col.val[0], vget_low_u16(b0_u16.val[0])), - vaddw_u16(sum_col.val[1], vget_high_u16(b0_u16.val[0])), - vaddw_u16(sum_col.val[2], vget_low_u16(b0_u16.val[1])), - vaddw_u16(sum_col.val[3], vget_high_u16(b0_u16.val[1])) + vaddw_s16(sum_col.val[0], vget_low_s16(b0_s16.val[0])), + vaddw_s16(sum_col.val[1], vget_high_s16(b0_s16.val[0])), + vaddw_s16(sum_col.val[2], vget_low_s16(b0_s16.val[1])), + vaddw_s16(sum_col.val[3], vget_high_s16(b0_s16.val[1])) } }; } auto vector_sum_col = reinterpret_cast(out.ptr()); - vst1q_s32(vector_sum_col + 0, vreinterpretq_s32_u32(sum_col.val[0])); - vst1q_s32(vector_sum_col + 4, vreinterpretq_s32_u32(sum_col.val[1])); - vst1q_s32(vector_sum_col + 8, vreinterpretq_s32_u32(sum_col.val[2])); - vst1q_s32(vector_sum_col + 12, vreinterpretq_s32_u32(sum_col.val[3])); + vst1q_s32(vector_sum_col + 0, sum_col.val[0]); + vst1q_s32(vector_sum_col + 4, sum_col.val[1]); + vst1q_s32(vector_sum_col + 8, sum_col.val[2]); + vst1q_s32(vector_sum_col + 12, sum_col.val[3]); }, in, out); } @@ -326,17 +326,17 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf } // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation - uint32x4x4_t sum_col = + int32x4x4_t sum_col = { { - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0), - vdupq_n_u32(0) + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0), + vdupq_n_s32(0) } }; - const uint8_t *matrix_b = inb.ptr() + id.y() * _input->info()->strides_in_bytes()[2]; + auto matrix_b = reinterpret_cast(inb.ptr() + id.y() * _input->info()->strides_in_bytes()[2]); #if __arm__ asm volatile("PLD [%0, #128*4]" ::"r"(matrix_b)); @@ -347,10 +347,10 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf // This for loop performs 4 accumulations for(; i <= (_k - 4); i += 4) { - const uint8x16_t b0_u8 = vld1q_u8(matrix_b + 0 * in_b_stride); - const uint8x16_t b1_u8 = vld1q_u8(matrix_b + 1 * in_b_stride); - const uint8x16_t b2_u8 = vld1q_u8(matrix_b + 2 * in_b_stride); - const uint8x16_t b3_u8 = vld1q_u8(matrix_b + 3 * in_b_stride); + const int8x16_t b0_s8 = vld1q_s8(matrix_b + 0 * in_b_stride); + const int8x16_t b1_s8 = vld1q_s8(matrix_b + 1 * in_b_stride); + const int8x16_t b2_s8 = vld1q_s8(matrix_b + 2 * in_b_stride); + const int8x16_t b3_s8 = vld1q_s8(matrix_b + 3 * in_b_stride); #if __arm__ asm volatile("PLD [%0, #128*1]" ::"r"(matrix_b + 1 * in_b_stride)); @@ -360,31 +360,31 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf #endif /* __arm__ */ // Partial accumulation in u16 - uint16x8x2_t tmp_sum = + int16x8x2_t tmp_sum = { { - vdupq_n_u16(0), - vdupq_n_u16(0) + vdupq_n_s16(0), + vdupq_n_s16(0) } }; - tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b0_u8)); - tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b1_u8)); - tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b2_u8)); - tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b3_u8)); - tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b0_u8)); - tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b1_u8)); - tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b2_u8)); - tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b3_u8)); + tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b0_s8)); + tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b1_s8)); + tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b2_s8)); + tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b3_s8)); + tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b0_s8)); + tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b1_s8)); + tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b2_s8)); + tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b3_s8)); // Accumulate to U32 sum_col = { { - vaddw_u16(sum_col.val[0], vget_low_u16(tmp_sum.val[0])), - vaddw_u16(sum_col.val[1], vget_high_u16(tmp_sum.val[0])), - vaddw_u16(sum_col.val[2], vget_low_u16(tmp_sum.val[1])), - vaddw_u16(sum_col.val[3], vget_high_u16(tmp_sum.val[1])) + vaddw_s16(sum_col.val[0], vget_low_s16(tmp_sum.val[0])), + vaddw_s16(sum_col.val[1], vget_high_s16(tmp_sum.val[0])), + vaddw_s16(sum_col.val[2], vget_low_s16(tmp_sum.val[1])), + vaddw_s16(sum_col.val[3], vget_high_s16(tmp_sum.val[1])) } }; @@ -394,14 +394,14 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf // This for loop perfoms the leftover accumulations for(; i < _k; ++i) { - const uint8x16_t b0_u8 = vld1q_u8(matrix_b + 0 * in_b_stride); + const int8x16_t b0_s8 = vld1q_s8(matrix_b + 0 * in_b_stride); - // Convert U8 to U16 - const uint16x8x2_t b0_u16 = + // Convert S8 to S16 + const int16x8x2_t b0_s16 = { { - vmovl_u8(vget_low_u8(b0_u8)), - vmovl_u8(vget_high_u8(b0_u8)) + vmovl_s8(vget_low_s8(b0_s8)), + vmovl_s8(vget_high_s8(b0_s8)) } }; @@ -409,10 +409,10 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf sum_col = { { - vaddw_u16(sum_col.val[0], vget_low_u16(b0_u16.val[0])), - vaddw_u16(sum_col.val[1], vget_high_u16(b0_u16.val[0])), - vaddw_u16(sum_col.val[2], vget_low_u16(b0_u16.val[1])), - vaddw_u16(sum_col.val[3], vget_high_u16(b0_u16.val[1])) + vaddw_s16(sum_col.val[0], vget_low_s16(b0_s16.val[0])), + vaddw_s16(sum_col.val[1], vget_high_s16(b0_s16.val[0])), + vaddw_s16(sum_col.val[2], vget_low_s16(b0_s16.val[1])), + vaddw_s16(sum_col.val[3], vget_high_s16(b0_s16.val[1])) } }; @@ -421,11 +421,11 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf auto vector_sum_col = reinterpret_cast(out.ptr()); - vst1q_s32(vector_sum_col + 0, vreinterpretq_s32_u32(sum_col.val[0])); - vst1q_s32(vector_sum_col + 4, vreinterpretq_s32_u32(sum_col.val[1])); - vst1q_s32(vector_sum_col + 8, vreinterpretq_s32_u32(sum_col.val[2])); - vst1q_s32(vector_sum_col + 12, vreinterpretq_s32_u32(sum_col.val[3])); + vst1q_s32(vector_sum_col + 0, sum_col.val[0]); + vst1q_s32(vector_sum_col + 4, sum_col.val[1]); + vst1q_s32(vector_sum_col + 8, sum_col.val[2]); + vst1q_s32(vector_sum_col + 12, sum_col.val[3]); }, inb, out); } -} \ No newline at end of file +} diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp index 8728e77d9e..5fe198f455 100644 --- a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp +++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp @@ -35,6 +35,12 @@ #include "arm_compute/core/Window.h" #include "support/ToolchainSupport.h" +namespace arm_compute +{ +#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_12x8.hpp" +} // namespace arm_compute + #include #include #include @@ -42,91 +48,22 @@ // Enable only if compiled for AArch64-V8.2-A targets #ifdef ARM_COMPUTE_AARCH64_V8_2 -#define ASM_PREFETCH(address) "PRFM PLDL1KEEP, " address "\n" -#define ASM_PREFETCHL2(address) "PRFM PLDL2KEEP, " address "\n" -#define ASM_PREFETCHW(address) "PRFM PSTL1KEEP, " address "\n" -#define ASM_PREFETCHWL2(address) "PRFM PSTL2KEEP, " address "\n" - -static inline void stincpld(uint32x4_t v0, uint32x4_t v1, uint32x4_t v2, uint32x4_t v3, - uint32x4_t v4, uint32x4_t v5, uint32x4_t v6, uint32x4_t v7, - uint32_t *&ptr0, uint32_t *&ptr1, uint32_t *&ptr2, uint32_t *&ptr3, - uint32_t *&ptr4, uint32_t *&ptr5, uint32_t *&ptr6, uint32_t *&ptr7) -{ - __asm __volatile( - "LDR q0, [%[ptr0]]\n" - "LDR q1, [%[ptr1]]\n" - "LDR q2, [%[ptr2]]\n" - "LDR q3, [%[ptr3]]\n" - "LDR q4, [%[ptr4]]\n" - "LDR q5, [%[ptr5]]\n" - "LDR q6, [%[ptr6]]\n" - "LDR q7, [%[ptr7]]\n" - "ADD v0.4s, v0.4s, %[v0].4s\n" ASM_PREFETCH("[%[ptr0], #80]") "ADD v1.4s, v1.4s, %[v1].4s\n" ASM_PREFETCH("[%[ptr1], #80]") "ADD v2.4s, v2.4s, %[v2].4s\n" ASM_PREFETCH("[%[ptr2], #80]") - "ADD v3.4s, v3.4s, %[v3].4s\n" ASM_PREFETCH("[%[ptr3], #80]") "ADD v4.4s, v4.4s, %[v4].4s\n" ASM_PREFETCH("[%[ptr4], #80]") "ADD v5.4s, v5.4s, %[v5].4s\n" ASM_PREFETCH("[%[ptr5], #80]") - "ADD v6.4s, v6.4s, %[v6].4s\n" ASM_PREFETCH("[%[ptr6], #80]") "ADD v7.4s, v7.4s, %[v7].4s\n" ASM_PREFETCH("[%[ptr7], #80]") - "STR q0, [%[ptr0]], #16\n" - "STR q1, [%[ptr1]], #16\n" - "STR q2, [%[ptr2]], #16\n" - "STR q3, [%[ptr3]], #16\n" - "STR q4, [%[ptr4]], #16\n" - "STR q5, [%[ptr5]], #16\n" - "STR q6, [%[ptr6]], #16\n" - "STR q7, [%[ptr7]], #16\n" - : [ptr0] "+r"(ptr0), [ptr1] "+r"(ptr1), [ptr2] "+r"(ptr2), [ptr3] "+r"(ptr3), - [ptr4] "+r"(ptr4), [ptr5] "+r"(ptr5), [ptr6] "+r"(ptr6), [ptr7] "+r"(ptr7) - : [v0] "w"(v0), [v1] "w"(v1), [v2] "w"(v2), [v3] "w"(v3), - [v4] "w"(v4), [v5] "w"(v5), [v6] "w"(v6), [v7] "w"(v7) - : "x20", "x21", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory"); -} - -static inline void stinc(uint32x4_t v0, uint32x4_t v1, uint32x4_t v2, uint32x4_t v3, - uint32x4_t v4, uint32x4_t v5, uint32x4_t v6, uint32x4_t v7, - uint32_t *&ptr0, uint32_t *&ptr1, uint32_t *&ptr2, uint32_t *&ptr3, - uint32_t *&ptr4, uint32_t *&ptr5, uint32_t *&ptr6, uint32_t *&ptr7) -{ - __asm __volatile( - "LDR q0, [%[ptr0]]\n" - "LDR q1, [%[ptr1]]\n" - "LDR q2, [%[ptr2]]\n" - "LDR q3, [%[ptr3]]\n" - "LDR q4, [%[ptr4]]\n" - "LDR q5, [%[ptr5]]\n" - "LDR q6, [%[ptr6]]\n" - "LDR q7, [%[ptr7]]\n" - "ADD v0.4s, v0.4s, %[v0].4s\n" - "ADD v1.4s, v1.4s, %[v1].4s\n" - "ADD v2.4s, v2.4s, %[v2].4s\n" - "ADD v3.4s, v3.4s, %[v3].4s\n" - "ADD v4.4s, v4.4s, %[v4].4s\n" - "ADD v5.4s, v5.4s, %[v5].4s\n" - "ADD v6.4s, v6.4s, %[v6].4s\n" - "ADD v7.4s, v7.4s, %[v7].4s\n" - "STR q0, [%[ptr0]], #16\n" - "STR q1, [%[ptr1]], #16\n" - "STR q2, [%[ptr2]], #16\n" - "STR q3, [%[ptr3]], #16\n" - "STR q4, [%[ptr4]], #16\n" - "STR q5, [%[ptr5]], #16\n" - "STR q6, [%[ptr6]], #16\n" - "STR q7, [%[ptr7]], #16\n" - : [ptr0] "+r"(ptr0), [ptr1] "+r"(ptr1), [ptr2] "+r"(ptr2), [ptr3] "+r"(ptr3), - [ptr4] "+r"(ptr4), [ptr5] "+r"(ptr5), [ptr6] "+r"(ptr6), [ptr7] "+r"(ptr7) - : [v0] "w"(v0), [v1] "w"(v1), [v2] "w"(v2), [v3] "w"(v3), - [v4] "w"(v4), [v5] "w"(v5), [v6] "w"(v6), [v7] "w"(v7) - : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory"); -} - namespace arm_compute { -void NEGEMMLowpAArch64V8P4Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output) +void NEGEMMLowpAArch64V8P4Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::U8); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); - _input0 = input0; - _input1 = input1; - _output = output; + _input0 = input0; + _input1 = input1; + _output = output; + _workspace = workspace; + _alpha = alpha; + _beta = beta; + _transform_0 = transform_0; + _transform_1 = transform_1; // Configure kernel window Window win = calculate_max_window(*output->info()); @@ -149,375 +86,49 @@ bool NEGEMMLowpAArch64V8P4Kernel::is_parallelisable() const return false; } -#define _UDOT_MACRO \ - ".altmacro\n" \ - ".macro udot opd:req, opn:req, opm:req\n" \ - "local vd, vn, vm, h, l\n" \ - ".irp reg,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31\n" \ - ".ifeqs \"\\opd\",\"v\\reg\\.4s\"\n" \ - ".set vd,\\reg\n" \ - ".endif\n" \ - ".ifeqs \"\\opn\",\"v\\reg\\.16b\"\n" \ - ".set vn,\\reg\n" \ - ".endif\n" \ - ".irp idx,0,1,2,3\n" \ - ".ifeqs \"\\opm\",\"v\\reg\\.4b[\\idx\\]\"\n" \ - ".set vm,\\reg\n" \ - ".set h,\\idx / 2\n" \ - ".set l,\\idx %% 2\n" \ - ".endif\n" \ - ".endr\n" \ - ".endr\n" \ - ".ifndef vd\n" \ - ".error \"Bad operand \\opd\"\n" \ - ".exitm\n" \ - ".endif\n" \ - ".ifndef vn\n" \ - ".error \"Bad operand \\opn\"\n" \ - ".exitm\n" \ - ".endif\n" \ - ".ifndef vm\n" \ - ".error \"Bad operand \\opm\"\n" \ - ".exitm\n" \ - ".endif\n" \ - ".ifndef h\n" \ - ".error \"Bad operand \\opm\"\n" \ - ".exitm\n" \ - ".endif\n" \ - ".ifndef l\n" \ - ".error \"Bad operand \\opm\"\n" \ - ".exitm\n" \ - ".endif\n" \ - ".int 0x6f80e000 | vd | (vn << 5) | (vm << 16) | (l << 21) | (h << 11)\n" \ - ".endm\n" - -#define _PREFETCH_ \ - __asm __volatile( \ - "" ASM_PREFETCH("[%[a_ptr], #64]") \ - ASM_PREFETCH("[%[a_ptr], #128]") \ - ASM_PREFETCH("[%[a_ptr], #192]") \ - : \ - : \ - [a_ptr] "r"(a_ptr), [b_ptr] "r"(b_ptr) \ - : "x20", "x21", "memory"); \ - __asm __volatile( \ - "" ASM_PREFETCH("[%[b_ptr]]") \ - ASM_PREFETCH("[%[b_ptr], #64]") \ - ASM_PREFETCH("[%[b_ptr], #128]") \ - ASM_PREFETCH("[%[b_ptr], #192]") \ - : \ - : \ - [b_ptr] "r"(b_ptr) \ - : "x20", "x21"); \ - __asm __volatile( \ - "" \ - : [r00] "+w"(r00), [r01] "+w"(r01), \ - [r10] "+w"(r10), [r11] "+w"(r11), \ - [r20] "+w"(r20), [r21] "+w"(r21), \ - [r30] "+w"(r30), [r31] "+w"(r31), \ - [a0] "+w"(a0), [a1] "+w"(a1), \ - [b0] "+w"(b0), [b1] "+w"(b1), [b2] "=w"(b2), \ - [a_ptr] "+r"(a_ptr), [b_ptr] "+r"(b_ptr) \ - : \ - :); \ - __asm __volatile( \ - "" \ - : [r02] "+w"(r02), \ - [r12] "+w"(r12), \ - [r22] "+w"(r22), \ - [r32] "+w"(r32), \ - [r40] "+w"(r40), \ - [r50] "+w"(r50), \ - [r60] "+w"(r60), \ - [r70] "+w"(r70), \ - [a0a] "=w"(a0a), [a1a] "=w"(a1a), \ - [b0] "+w"(b0), [b2] "+w"(b2), [b5] "=&w"(b5) \ - : \ - :); \ - __asm __volatile( \ - "" \ - : \ - [r41] "+w"(r41), [r42] "+w"(r42), \ - [r51] "+w"(r51), [r52] "+w"(r52), \ - [r61] "+w"(r61), [r62] "+w"(r62), \ - [r71] "+w"(r71), [r72] "+w"(r72), \ - [a1] "+w"(a1), \ - [b0] "+w"(b0), [b1] "+w"(b1), [b2] "+w"(b2), \ - [b_ptr] "+r"(b_ptr), [k] "+r"(k) \ - : \ - :); - void NEGEMMLowpAArch64V8P4Kernel::run(const Window &window, const ThreadInfo &info) { - ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - const int x_block = 348; - const int k_block = 1664; - const int nthreads = 1; - const int M = _output->info()->tensor_shape().y(); - const int N = _output->info()->tensor_shape().x(); - const int K = _input0->info()->tensor_shape().x() >> 3; - - int yblocksperthread = ((M / nthreads) + 7) / 8; - - if(yblocksperthread < 1) - { - yblocksperthread = 1; - } - - const int lda = _input0->info()->strides_in_bytes().y(); - const int ldb = _input1->info()->strides_in_bytes().y(); - const int ldc = _output->info()->strides_in_bytes().y(); - const int ldc2 = _output->info()->strides_in_bytes().x(); - const int ldc3 = ldc / sizeof(uint32_t); + const int lda = _input0->info()->strides_in_bytes().y(); + const int ldb = _input1->info()->strides_in_bytes().y(); + const int ldc = _output->info()->strides_in_bytes().y() / sizeof(int32_t); - const int threadid = 0; - int y0 = threadid * yblocksperthread * 8; - int ymax = (threadid + 1) * yblocksperthread * 8; - if(y0 >= M) - { - return; - } - if(ymax > M) - { - ymax = M; - } - for(int k0 = 0; k0 < K; k0 += k_block) - { - int kmax = k0 + k_block; - if(kmax > K) - { - kmax = K; - } + const auto in1_ptr = reinterpret_cast(_input1->buffer()); - for(int x0 = 0; x0 < N; x0 += x_block) - { - int xmax = x0 + x_block; - if(xmax > N) - { - xmax = N; - } + const int M = std::min(_output->info()->tensor_shape().y(), static_cast(window.y().end())) - window.y().start(); + const int N = _output->info()->tensor_shape().x(); + const int K = _input0->info()->tensor_shape().x(); - for(int y = y0; y < ymax; y += 8) - { - auto c_ptr0 = reinterpret_cast(_output->buffer() + (y * ldc) + x0 * ldc2); - uint32_t *c_ptr1 = c_ptr0 + ldc3; - uint32_t *c_ptr2 = c_ptr1 + ldc3; - uint32_t *c_ptr3 = c_ptr2 + ldc3; - uint32_t *c_ptr4 = c_ptr3 + ldc3; - uint32_t *c_ptr5 = c_ptr4 + ldc3; - uint32_t *c_ptr6 = c_ptr5 + ldc3; - uint32_t *c_ptr7 = c_ptr6 + ldc3; + // Only iterate over batches + Window win(window); + win.set(0, Window::Dimension(0, 1, 1)); + win.set(1, Window::Dimension(0, 1, 1)); - __asm __volatile( - "" ASM_PREFETCH("[%[c_ptr0]]") - ASM_PREFETCH("[%[c_ptr1]]") - ASM_PREFETCH("[%[c_ptr2]]") - ASM_PREFETCH("[%[c_ptr3]]") - ASM_PREFETCH("[%[c_ptr4]]") - ASM_PREFETCH("[%[c_ptr5]]") - ASM_PREFETCH("[%[c_ptr6]]") - ASM_PREFETCH("[%[c_ptr7]]") - : - : [c_ptr0] "r"(c_ptr0), [c_ptr1] "r"(c_ptr1), [c_ptr2] "r"(c_ptr2), [c_ptr3] "r"(c_ptr3), - [c_ptr4] "r"(c_ptr4), [c_ptr5] "r"(c_ptr5), [c_ptr6] "r"(c_ptr6), [c_ptr7] "r"(c_ptr7) - : "x20", "x21"); + Iterator in0(_input0, window); + Iterator out(_output, window); - for(int x = x0; x < xmax; x += 12) - { - register uint32x4_t r00 asm("v8"); - register uint32x4_t r10 asm("v9"); - register uint32x4_t r20 asm("v10"); - register uint32x4_t r30 asm("v11"); - register uint32x4_t r40 asm("v12"); - register uint32x4_t r50 asm("v13"); - register uint32x4_t r60 asm("v14"); - register uint32x4_t r70 asm("v15"); - register uint32x4_t r01 asm("v16"); - register uint32x4_t r11 asm("v17"); - register uint32x4_t r21 asm("v18"); - register uint32x4_t r31 asm("v19"); - register uint32x4_t r41 asm("v20"); - register uint32x4_t r51 asm("v21"); - register uint32x4_t r61 asm("v22"); - register uint32x4_t r71 asm("v23"); - register uint32x4_t r02 asm("v24"); - register uint32x4_t r12 asm("v25"); - register uint32x4_t r22 asm("v26"); - register uint32x4_t r32 asm("v27"); - register uint32x4_t r42 asm("v28"); - register uint32x4_t r52 asm("v29"); - register uint32x4_t r62 asm("v30"); - register uint32x4_t r72 asm("v31"); + GemmInterleaved gemm(&info.cpu_info, M, N, K, !_transform_1, !_transform_1); - register uint8x16_t a0 asm("v0"); - register uint8x16_t a1 asm("v1"); - register uint8x16_t b0 asm("v2"); - register uint8x16_t b1 asm("v3"); - register uint8x16_t b2 asm("v4"); - register uint8x16_t a0a asm("v5"); - register uint8x16_t a1a asm("v6"); - register uint8x16_t b5 asm("v7"); - const uint8_t *a_ptr = _input0->buffer() + ((y / 8) * lda) + (k0 * 8); - const uint8_t *b_ptr = _input1->buffer() + ((x / 12) * ldb) + (k0 * 12); + constexpr size_t alignment = 4096; + const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id; + void *workspace = _workspace->buffer() + offset; + size_t workspace_size = _workspace->info()->total_size(); - r00 = r01 = r02 = r10 = r11 = r12 = r20 = r21 = r22 = r30 = r31 = r32 = vdupq_n_u32(0); - r40 = r41 = r42 = r50 = r51 = r52 = r60 = r61 = r62 = r70 = r71 = r72 = vdupq_n_u32(0); - - int k = ((kmax - k0) / 8) - 1; - - a0 = vld1q_u8(a_ptr); - b0 = vld1q_u8(b_ptr); - a1 = vld1q_u8(a_ptr + 16); - b1 = vld1q_u8(b_ptr + 16); - - _PREFETCH_ - - __asm __volatile( - _UDOT_MACRO - "1:\n" - "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" - "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - "ldr %q[b2], [%[b_ptr], #32]\n" - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %q[a0a], [%[a_ptr], #32]\n" - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" - "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" - "ldr %q[a1a], [%[a_ptr], #48]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #48]\n" - - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" ASM_PREFETCH("[%[a_ptr], #256]") - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #64]\n" - - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" ASM_PREFETCH("[%[b_ptr], #256]") - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "ldr %q[b2], [%[b_ptr], #80]\n" - - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" - "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" - "ldr %q[a0], [%[a_ptr], #64]\n" - "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" - "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" - "ldr %q[a1], [%[a_ptr], #80]\n" - "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" - "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #96]\n" - - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" ASM_PREFETCH("[%[b_ptr], #320]") - "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" - "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" - "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #112]\n" - - "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" - "add %[a_ptr], %[a_ptr], #64\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "add %[b_ptr], %[b_ptr], #96\n" - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" - "subs %w[k], %w[k], #1\n" - "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" - - "bne 1b\n" - - "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" - "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - "ldr %q[b2], [%[b_ptr], #32]\n" - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %q[a0a], [%[a_ptr], #32]\n" - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" - "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" - "ldr %q[a1a], [%[a_ptr], #48]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #48]\n" - - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #64]\n" - - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "ldr %q[b2], [%[b_ptr], #80]\n" - - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" ASM_PREFETCH("[%[c_ptr0]]") "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" ASM_PREFETCH("[%[c_ptr1]]") "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" - ASM_PREFETCH("[%[c_ptr2]]") "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" ASM_PREFETCH("[%[c_ptr3]]") "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" ASM_PREFETCH("[%[c_ptr4]]") - "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" ASM_PREFETCH("[%[c_ptr5]]") "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" ASM_PREFETCH("[%[c_ptr6]]") "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - ASM_PREFETCH("[%[c_ptr7]]") - - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" ASM_PREFETCH("[%[c_ptr0], #48]") "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" ASM_PREFETCH("[%[c_ptr1], #48]") "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" - ASM_PREFETCH("[%[c_ptr2], #48]") "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" ASM_PREFETCH("[%[c_ptr3], #48]") "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" ASM_PREFETCH("[%[c_ptr4], #48]") - "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" ASM_PREFETCH("[%[c_ptr5], #48]") "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" ASM_PREFETCH("[%[c_ptr6], #48]") "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" - ASM_PREFETCH("[%[c_ptr7], #48]") - - "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "add %[b_ptr], %[b_ptr], #96\n" - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" - "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" - - // Clean up macro namespace - ".purgem udot\n" - - : - [a_ptr] "+r"(a_ptr), [b_ptr] "+r"(b_ptr), - [a0] "+w"(a0), [a1] "+w"(a1), [a0a] "+w"(a0a), [a1a] "+w"(a1a), - [b0] "+w"(b0), [b1] "+w"(b1), [b2] "+w"(b2), [k] "+r"(k) - : [c_ptr0] "r"(c_ptr0), [c_ptr1] "r"(c_ptr1), [c_ptr2] "r"(c_ptr2), [c_ptr3] "r"(c_ptr3), - [c_ptr4] "r"(c_ptr4), [c_ptr5] "r"(c_ptr5), [c_ptr6] "r"(c_ptr6), [c_ptr7] "r"(c_ptr7) - : "x20", "x21"); - - stincpld(r00, r10, r20, r30, r40, r50, r60, r70, c_ptr0, c_ptr1, c_ptr2, c_ptr3, c_ptr4, c_ptr5, c_ptr6, c_ptr7); - stinc(r01, r11, r21, r31, r41, r51, r61, r71, c_ptr0, c_ptr1, c_ptr2, c_ptr3, c_ptr4, c_ptr5, c_ptr6, c_ptr7); - stinc(r02, r12, r22, r32, r42, r52, r62, r72, c_ptr0, c_ptr1, c_ptr2, c_ptr3, c_ptr4, c_ptr5, c_ptr6, c_ptr7); - } - } - } + if(support::cpp11::align(alignment, gemm.get_working_size(), workspace, workspace_size) == nullptr) + { + ARM_COMPUTE_ERROR("Not enough space to align buffer!"); } + + execute_window_loop(win, [&](const Coordinates & id) + { + gemm.execute(reinterpret_cast(in0.ptr()), lda, + reinterpret_cast(in1_ptr), ldb, + reinterpret_cast(out.ptr()), ldc, + _alpha, _beta, workspace); + }, + in0, out); } } // namespace arm_compute #endif /* ARM_COMPUTE_AARCH64_V8_2 */ -- cgit v1.2.1