From 30eacad2905e375540fa800dccb4d62b053d3df5 Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Sat, 11 Jan 2025 17:14:22 -0500 Subject: [PATCH] use calc_superblock everywhere --- .../vulkan-shaders/mul_mat_vec_q4_k.comp | 161 +++++++------ .../vulkan-shaders/mul_mat_vec_q5_k.comp | 225 +++++++++--------- 2 files changed, 196 insertions(+), 190 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp index dc1540a5b..f9cde0648 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp @@ -6,6 +6,86 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y1_idx = i * QUANT_K + y_offset; + const uint y2_idx = y1_idx + 128; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + vec2 d = vec2(data_a[ib0 + i].d); + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + + const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; + const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; + const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); + const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); + + const FLOAT_TYPE sc0 = scale_0_4_l_f.x; + const FLOAT_TYPE sc1 = scale_0_4_l_f.y; + const FLOAT_TYPE sc2 = scale_0_4_l_f.z; + const FLOAT_TYPE sc3 = scale_0_4_l_f.w; + const FLOAT_TYPE sc4 = scale8_f.x; + const FLOAT_TYPE sc5 = scale8_f.y; + const FLOAT_TYPE sc6 = scale8_f.z; + const FLOAT_TYPE sc7 = scale8_f.w; + + const uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; + const uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; + + const uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; + const uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; + const uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F; + const uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F; + + const vec4 qs0_lo4 = vec4(unpack8(qs0_u32_lo4)); + const vec4 qs64_lo4 = vec4(unpack8(qs64_u32_lo4)); + const vec4 qs0_hi4 = vec4(unpack8(qs0_u32_hi4)); + const vec4 qs64_hi4 = vec4(unpack8(qs64_u32_hi4)); + + const FLOAT_TYPE q4_0 = qs0_lo4.x; + const FLOAT_TYPE q4_1 = qs0_lo4.y; + const FLOAT_TYPE q4_2 = qs0_lo4.z; + const FLOAT_TYPE q4_3 = qs0_lo4.w; + const FLOAT_TYPE q4_4 = qs0_hi4.x; + const FLOAT_TYPE q4_5 = qs0_hi4.y; + const FLOAT_TYPE q4_6 = qs0_hi4.z; + const FLOAT_TYPE q4_7 = qs0_hi4.w; + const FLOAT_TYPE q4_8 = qs64_lo4.x; + const FLOAT_TYPE q4_9 = qs64_lo4.y; + const FLOAT_TYPE q4_10 = qs64_lo4.z; + const FLOAT_TYPE q4_11 = qs64_lo4.w; + const FLOAT_TYPE q4_12 = qs64_hi4.x; + const FLOAT_TYPE q4_13 = qs64_hi4.y; + const FLOAT_TYPE q4_14 = qs64_hi4.z; + const FLOAT_TYPE q4_15 = qs64_hi4.w; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 by10 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 ]); + vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]); + vec4 by20 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 ]); + vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]); + + const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3))); + const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7))); + const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11))); + const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15))); + const FLOAT_TYPE smin = + fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7, + fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7, + fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7, + fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7))))))))))))))); + temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); + } + } +} + void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -29,91 +109,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint q_offset = 32*v_im + l0; const uint y_offset = 64*v_im + l0; - FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; - [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { temp[j][i] = FLOAT_TYPE(0); } } - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) { - const uint y1_idx = i * QUANT_K + y_offset; - const uint y2_idx = y1_idx + 128; - - [[unroll]] for (uint n = 0; n < num_rows; ++n) { - const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; - vec2 d = vec2(data_a[ib0 + i].d); - const FLOAT_TYPE dall = FLOAT_TYPE(d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - - const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; - const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; - const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; - - const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; - const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; - const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); - const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); - - const FLOAT_TYPE sc0 = scale_0_4_l_f.x; - const FLOAT_TYPE sc1 = scale_0_4_l_f.y; - const FLOAT_TYPE sc2 = scale_0_4_l_f.z; - const FLOAT_TYPE sc3 = scale_0_4_l_f.w; - const FLOAT_TYPE sc4 = scale8_f.x; - const FLOAT_TYPE sc5 = scale8_f.y; - const FLOAT_TYPE sc6 = scale8_f.z; - const FLOAT_TYPE sc7 = scale8_f.w; - - const uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; - const uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; - - const uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; - const uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; - const uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F; - const uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F; - - const vec4 qs0_lo4 = vec4(unpack8(qs0_u32_lo4)); - const vec4 qs64_lo4 = vec4(unpack8(qs64_u32_lo4)); - const vec4 qs0_hi4 = vec4(unpack8(qs0_u32_hi4)); - const vec4 qs64_hi4 = vec4(unpack8(qs64_u32_hi4)); - - const FLOAT_TYPE q4_0 = qs0_lo4.x; - const FLOAT_TYPE q4_1 = qs0_lo4.y; - const FLOAT_TYPE q4_2 = qs0_lo4.z; - const FLOAT_TYPE q4_3 = qs0_lo4.w; - const FLOAT_TYPE q4_4 = qs0_hi4.x; - const FLOAT_TYPE q4_5 = qs0_hi4.y; - const FLOAT_TYPE q4_6 = qs0_hi4.z; - const FLOAT_TYPE q4_7 = qs0_hi4.w; - const FLOAT_TYPE q4_8 = qs64_lo4.x; - const FLOAT_TYPE q4_9 = qs64_lo4.y; - const FLOAT_TYPE q4_10 = qs64_lo4.z; - const FLOAT_TYPE q4_11 = qs64_lo4.w; - const FLOAT_TYPE q4_12 = qs64_hi4.x; - const FLOAT_TYPE q4_13 = qs64_hi4.y; - const FLOAT_TYPE q4_14 = qs64_hi4.z; - const FLOAT_TYPE q4_15 = qs64_hi4.w; - - [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { - vec4 by10 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 ]); - vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]); - vec4 by20 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 ]); - vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]); - - const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3))); - const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7))); - const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11))); - const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15))); - const FLOAT_TYPE smin = - fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7, - fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7, - fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7, - fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7))))))))))))))); - temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); - } - } - } + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) + calc_superblock(a_offset, b_offset, v_im, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows); reduce_result(temp, d_offset, first_row, num_rows, tid); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp index df2a8bd20..6c84ef3cd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp @@ -6,6 +6,118 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint l0, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y1_idx = i * QUANT_K + y_offset; + const uint y2_idx = y1_idx + 128; + + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + vec2 d = vec2(data_a[ib0 + i].d); + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + + const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; + const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; + const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); + const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); + + const FLOAT_TYPE sc0 = scale_0_4_l_f.x; + const FLOAT_TYPE sc1 = scale_0_4_l_f.y; + const FLOAT_TYPE sc2 = scale_0_4_l_f.z; + const FLOAT_TYPE sc3 = scale_0_4_l_f.w; + const FLOAT_TYPE sc4 = scale8_f.x; + const FLOAT_TYPE sc5 = scale8_f.y; + const FLOAT_TYPE sc6 = scale8_f.z; + const FLOAT_TYPE sc7 = scale8_f.w; + + const uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); + const uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16); + + uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F; + uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F; + uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F; + uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F; + + const uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8])); + + const uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4; + const uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3; + const uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010); + const uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1; + + qs0_16_u32_lo4 += qs0_16_lo4_offset16; + qs0_16_u32_hi4 += qs0_16_hi4_offset16; + qs64_80_u32_lo4 += qs64_80_lo4_offset16; + qs64_80_u32_hi4 += qs64_80_hi4_offset16; + + const vec4 qs0_16_lo4 = vec4(unpack8(qs0_16_u32_lo4)); + const vec4 qs64_80_lo4 = vec4(unpack8(qs64_80_u32_lo4)); + const vec4 qs0_16_hi4 = vec4(unpack8(qs0_16_u32_hi4)); + const vec4 qs64_80_hi4 = vec4(unpack8(qs64_80_u32_hi4)); + + const FLOAT_TYPE q4_0 = qs0_16_lo4.x; + const FLOAT_TYPE q4_1 = qs0_16_lo4.y; + const FLOAT_TYPE q4_2 = qs0_16_lo4.z; + const FLOAT_TYPE q4_3 = qs0_16_lo4.w; + const FLOAT_TYPE q4_4 = qs0_16_hi4.x; + const FLOAT_TYPE q4_5 = qs0_16_hi4.y; + const FLOAT_TYPE q4_6 = qs0_16_hi4.z; + const FLOAT_TYPE q4_7 = qs0_16_hi4.w; + const FLOAT_TYPE q4_8 = qs64_80_lo4.x; + const FLOAT_TYPE q4_9 = qs64_80_lo4.y; + const FLOAT_TYPE q4_10 = qs64_80_lo4.z; + const FLOAT_TYPE q4_11 = qs64_80_lo4.w; + const FLOAT_TYPE q4_12 = qs64_80_hi4.x; + const FLOAT_TYPE q4_13 = qs64_80_hi4.y; + const FLOAT_TYPE q4_14 = qs64_80_hi4.z; + const FLOAT_TYPE q4_15 = qs64_80_hi4.w; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec2 by10 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 ]); + vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8]); + vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]); + vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]); + vec2 by20 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 ]); + vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8]); + vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]); + vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]); + + const FLOAT_TYPE sx = + fma(FLOAT_TYPE(by10.x), q4_0, + fma(FLOAT_TYPE(by10.y), q4_1, + fma(FLOAT_TYPE(by116.x), q4_2, + FLOAT_TYPE(by116.y) * q4_3))); + const FLOAT_TYPE sy = + fma(FLOAT_TYPE(by132.x), q4_4, + fma(FLOAT_TYPE(by132.y), q4_5, + fma(FLOAT_TYPE(by148.x), q4_6, + FLOAT_TYPE(by148.y) * q4_7))); + const FLOAT_TYPE sz = + fma(FLOAT_TYPE(by20.x), q4_8, + fma(FLOAT_TYPE(by20.y), q4_9, + fma(FLOAT_TYPE(by216.x), q4_10, + FLOAT_TYPE(by216.y) * q4_11))); + const FLOAT_TYPE sw = + fma(FLOAT_TYPE(by232.x), q4_12, + fma(FLOAT_TYPE(by232.y), q4_13, + fma(FLOAT_TYPE(by248.x), q4_14, + FLOAT_TYPE(by248.y) * q4_15))); + const FLOAT_TYPE smin = + fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2, + fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3, + fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6, + (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7))); + temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); + } + } +} + void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -28,123 +140,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint q_offset = 32*v_im + l0; const uint y_offset = 64*v_im + l0; - FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; - [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { temp[j][i] = FLOAT_TYPE(0); } } - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) { - const uint y1_idx = i * QUANT_K + y_offset; - const uint y2_idx = y1_idx + 128; - - [[unroll]] for (uint n = 0; n < num_rows; ++n) { - const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; - vec2 d = vec2(data_a[ib0 + i].d); - const FLOAT_TYPE dall = FLOAT_TYPE(d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - - const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; - const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; - const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; - - const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; - const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2; - const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F)); - const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h)); - - const FLOAT_TYPE sc0 = scale_0_4_l_f.x; - const FLOAT_TYPE sc1 = scale_0_4_l_f.y; - const FLOAT_TYPE sc2 = scale_0_4_l_f.z; - const FLOAT_TYPE sc3 = scale_0_4_l_f.w; - const FLOAT_TYPE sc4 = scale8_f.x; - const FLOAT_TYPE sc5 = scale8_f.y; - const FLOAT_TYPE sc6 = scale8_f.z; - const FLOAT_TYPE sc7 = scale8_f.w; - - const uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); - const uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16); - - uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F; - uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F; - uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F; - uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F; - - const uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8])); - - const uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4; - const uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3; - const uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010); - const uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1; - - qs0_16_u32_lo4 += qs0_16_lo4_offset16; - qs0_16_u32_hi4 += qs0_16_hi4_offset16; - qs64_80_u32_lo4 += qs64_80_lo4_offset16; - qs64_80_u32_hi4 += qs64_80_hi4_offset16; - - const vec4 qs0_16_lo4 = vec4(unpack8(qs0_16_u32_lo4)); - const vec4 qs64_80_lo4 = vec4(unpack8(qs64_80_u32_lo4)); - const vec4 qs0_16_hi4 = vec4(unpack8(qs0_16_u32_hi4)); - const vec4 qs64_80_hi4 = vec4(unpack8(qs64_80_u32_hi4)); - - const FLOAT_TYPE q4_0 = qs0_16_lo4.x; - const FLOAT_TYPE q4_1 = qs0_16_lo4.y; - const FLOAT_TYPE q4_2 = qs0_16_lo4.z; - const FLOAT_TYPE q4_3 = qs0_16_lo4.w; - const FLOAT_TYPE q4_4 = qs0_16_hi4.x; - const FLOAT_TYPE q4_5 = qs0_16_hi4.y; - const FLOAT_TYPE q4_6 = qs0_16_hi4.z; - const FLOAT_TYPE q4_7 = qs0_16_hi4.w; - const FLOAT_TYPE q4_8 = qs64_80_lo4.x; - const FLOAT_TYPE q4_9 = qs64_80_lo4.y; - const FLOAT_TYPE q4_10 = qs64_80_lo4.z; - const FLOAT_TYPE q4_11 = qs64_80_lo4.w; - const FLOAT_TYPE q4_12 = qs64_80_hi4.x; - const FLOAT_TYPE q4_13 = qs64_80_hi4.y; - const FLOAT_TYPE q4_14 = qs64_80_hi4.z; - const FLOAT_TYPE q4_15 = qs64_80_hi4.w; - - [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { - vec2 by10 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 ]); - vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8]); - vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]); - vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]); - vec2 by20 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 ]); - vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8]); - vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]); - vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]); - - const FLOAT_TYPE sx = - fma(FLOAT_TYPE(by10.x), q4_0, - fma(FLOAT_TYPE(by10.y), q4_1, - fma(FLOAT_TYPE(by116.x), q4_2, - FLOAT_TYPE(by116.y) * q4_3))); - const FLOAT_TYPE sy = - fma(FLOAT_TYPE(by132.x), q4_4, - fma(FLOAT_TYPE(by132.y), q4_5, - fma(FLOAT_TYPE(by148.x), q4_6, - FLOAT_TYPE(by148.y) * q4_7))); - const FLOAT_TYPE sz = - fma(FLOAT_TYPE(by20.x), q4_8, - fma(FLOAT_TYPE(by20.y), q4_9, - fma(FLOAT_TYPE(by216.x), q4_10, - FLOAT_TYPE(by216.y) * q4_11))); - const FLOAT_TYPE sw = - fma(FLOAT_TYPE(by232.x), q4_12, - fma(FLOAT_TYPE(by232.y), q4_13, - fma(FLOAT_TYPE(by248.x), q4_14, - FLOAT_TYPE(by248.y) * q4_15))); - const FLOAT_TYPE smin = - fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2, - fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3, - fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6, - (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7))); - temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n])); - } - } - } + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) + calc_superblock(a_offset, b_offset, v_im, l0, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows); reduce_result(temp, d_offset, first_row, num_rows, tid); }